The ATLAS ROBIN – A High-Performance Data

The ATLAS ROBIN – A High-Performance
Data-Acquisition Module
Inauguraldissertation
zur Erlangung des akademischen Grades
eines Doktors der Naturwissenschaften
der Universität Mannheim
Vorgelegt von
Dipl. Ing. (FH) Andreas Kugel
aus Lindau
Mannheim, Juni 2009
Dekan:
Referent:
Korreferent:
Prof. Dr. Felix Freiling, Universität Mannheim
Prof. Dr. Reinhard Männer, Universität Mannheim
Prof. Dr. Peter Fischer, Universität Mannheim
Tag der mündlichen Prüfung: 19. August 2009
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Abstract
This work presents the re-configurable processor ROBIN, which is a key element of the dataacquisition-system of the ATLAS experiment, located at the new LHC at CERN. The ATLAS detector
provides data over 1600 channels simultaneously towards the DAQ system. The ATLAS dataflow
model follows the “PULL” strategy in contrast to the commonly used “PUSH” strategy. The data
volume transported is reduced by a factor of 10, however the data must be temporarily stored at the
entry to the DAQ system. The input layer consists of approx. 160 ROS read-out units comprising 1 PC
and 4 ROBIN modules. Each ROBIN device acquires detector data via 3 input channels and performs
local buffering. Board control is done via a 64-bit PCI interface. Event selection and data transmission
runs via PCI in the baseline bus-based ROS. Alternatively, a local GE interface can take over part or
all of the data traffic in the switch-based ROS, in order to reduce the load on the host PC. The
performance of the ROBIN module stems from the close cooperation of a fast embedded processor
with a complex FPGA. The efficient task-distribution lets the processor handle all complex
management functionality, programmed in “C” while all movement of data is performed by the FPGA
via multiple, concurrently operating DMA engines.
The ROBIN-project was carried-out by and international team and comprises the design specification,
the development of the ROBIN hardware, firmware (VHDL and C-Code), host-code (C++),
prototyping, volume production and installation of 700 boards. The project was led by the author of
this thesis. The hardware platform is an evolution of a FPGA processor previously designed by the
author. He has contributed elementary concepts of the communication mechanisms and the “C”-coded
embedded application software. He also organised and supervised the prototype and series productions
including the various design reports and presentations.
The results show that the ROBIN-module is able to meet its ambitious requirements of 100kHz
incoming fragment rate per channel with a concurrent outgoing fragment rate of 21kHz per channel.
At the system level, each ROS unit (12 channels) operates at the same rates, however for a subset of
the channels only. The ATLAS DAQ system – with 640 ROBIN modules installed – has performed a
successful data-taking phase at the start-up of the LHC in September.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
1
2
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Der ATLAS ROBIN – Eine Hochleistungs-Datenerfassungsbaugruppe
Zusammenfassung
Diese Arbeit beschreibt den re-konfigurierbaren Prozessor ROBIN, der ein Schlüsselelement im
Datenerfassungssystem des ATLAS-Experiments des LHC am CERN ist. Der ATLAS Detektor liefert
Daten über 1600 Kanäle gleichzeitig an das DAQ System. Im Gegensatz zur üblichen „PUSH“
Strategie für den Datentransport kommt bei ATLAS eine „PULL“ Strategie zur Anwendung, wodurch
das zu transportierende Datenvolumen um den Faktor 10 reduziert wird. Dazu müssen die Daten am
Eingang des DAQ System zwischengespeichert werden. Die Eingangsstufe nimmt die Daten in 160
ROS Ausleseeinheiten entgegen, die jeweils aus 1 PC mit 4 ROBIN Karten bestehen, jede ROBIN
Karte ist wiederum mit 3 Detektorkanälen verbunden. Die Daten werden auf den ROBINs gespeichert.
Die Überwachung der Baugruppe geschieht vom PC aus über ein 64-bit PCI Interface. In der Busbasierten Basisimplementierung des ROS erfolgt die Auswahl und Übertragung der Daten ebenfalls
über das PCI Interface. Eine lokale Gigabit-Ethernet Schnittstelle übernimmt in der alternativen
Netzwerk-basierten Implementierung einen Teil oder den gesamten Datenverkehr, um den PC zu
entlasten. Die hohe Leistungsfähigkeit und Flexibilität des ROBIN ergibt sich aus der Kombination
eines schnellen eingebetteten Prozessors mit einem hoch integrierten FPGA. In einer effektiven
Aufgabenverteilung bearbeitet der Prozessor alle komplexen Verwaltungsaufgaben, die in „C“
programmiert sind, während das FPGA sämtliche Datenbewegungen mit mehreren, gleichzeitig
arbeitenden DMA Einheiten durchführt.
Das ROBIN Projekt wurde von einem internationalen Team durchgeführt. Es umfasst die
Spezifikation des Designs, die Entwicklung der Hardware und der Firmware (VHDL und C) der
ROBIN Baugruppe, PC-Software (in C++) für Ansteuerung, Emulation und Test, Produktion und Test
von Prototypen sowie die Serienfertigung und Inbetriebnahme von 700 Baugruppen.
Das Projekt wurde unter Leitung des Autors dieser Arbeit durchgeführt. Die Hardwareplattform ist
eine Weiterentwicklung eines vom Autor entworfenen FPGA Prozessors. Grundlegende Konzepte der
Kommunikationsmechanismen stammen vom Autor, ebenso die „C“-Anwendungssoftware. Ebenfalls
wurde die Herstellung der Prototypen und die Serienfertigung inklusive der notwendigen
Statusberichte vom Autor vorbereitet und überwacht bzw. durchgeführt.
Die Ergebnisse zeigen, das die ROBIN Baugruppe die hohen Leistungsanforderungen von 100kHz
Ereignisrate am Eingang bei gleichzeitig 21kHz Abfragerate – jeweils auf allen 3 Kanälen gleichzeitig
– erfüllt. Auf Systemebene liegt der Arbeitspunkte für jede normale ROS-Einheit mit 12 Kanälen bei
100kHz Eingangsrate und 21kHz Abfragerate, jedoch nur für einen Teil der Kanäle. Das ATLAS DAQ
System mit 640 installierten ROBIN Baugruppen hat den Datenerfassungsbetrieb zum Start des LHC
erfolgreich aufgenommen.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
3
4
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Contents
1 Introduction......................................................................................................................................... 7
2 Data Acquisition Systems.................................................................................................................. 11
2.1 LHC Experiments..................................................................................................................... 11
2.1.1 CMS................................................................................................................................. 13
2.1.2 ALICE.............................................................................................................................. 17
2.2 LHCb........................................................................................................................................18
2.3 Summary.................................................................................................................................. 19
3 ATLAS DAQ..................................................................................................................................... 21
3.1 System View.............................................................................................................................21
3.2 Dataflow...................................................................................................................................25
3.3 ROS..........................................................................................................................................30
3.3.1 Event handling..................................................................................................................31
3.3.2 Configuration and monitoring...........................................................................................31
3.3.3 Baseline bus-based ROS...................................................................................................32
3.3.4 Switch-based ROS............................................................................................................37
3.4 Summary.................................................................................................................................. 38
4 FPGA Technology............................................................................................................................. 41
4.1 Device types............................................................................................................................. 41
4.2 Basic elements..........................................................................................................................41
4.3 Tools.........................................................................................................................................44
4.3.1 Specification.....................................................................................................................44
4.3.2 Simulation and test........................................................................................................... 45
4.3.3 Synthesis...........................................................................................................................47
4.3.4 Vendor specific tools.........................................................................................................48
4.4 IP-Cores....................................................................................................................................48
4.4.1 Embedded processors....................................................................................................... 49
4.4.2 Ethernet MAC.................................................................................................................. 49
4.5 Summary.................................................................................................................................. 50
5 ROBIN.............................................................................................................................................. 53
5.1 Requirements............................................................................................................................53
5.2 Project Management.................................................................................................................54
5.3 Implementation.........................................................................................................................55
5.3.1 Hardware.......................................................................................................................... 56
5.3.2 VHDL Firmware...............................................................................................................72
5.3.3 Software........................................................................................................................... 79
5.4 Installation and commissioning................................................................................................ 89
5.5 Summary.................................................................................................................................. 90
6 Results...............................................................................................................................................93
6.1 Stand-alone...............................................................................................................................93
6.1.1 PCI access.........................................................................................................................94
6.1.2 Network access.................................................................................................................96
6.1.3 Event bookkeeping........................................................................................................... 97
6.2 System performance................................................................................................................. 99
6.2.1 Bus-based ROS.................................................................................................................99
6.2.2 ATLAS Dataflow............................................................................................................ 100
6.3 Issues......................................................................................................................................101
6.4 Reliability............................................................................................................................... 103
6.5 Summary................................................................................................................................ 105
7 Discussion and Conclusions............................................................................................................ 107
7.1 Performance Assessment........................................................................................................ 107
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
5
7.2 Prospects.................................................................................................................................108
7.2.1 Short term developments................................................................................................ 108
7.2.2 Upgrade Phase-1.............................................................................................................109
7.2.3 Upgrade Phase-2.............................................................................................................110
7.3 Summary.................................................................................................................................110
8 Appendix..........................................................................................................................................111
8.1 Pre-ROBIN hardware..............................................................................................................111
8.1.1 MPRACE-1.....................................................................................................................111
8.1.2 µEnable...........................................................................................................................112
8.2 ROBIN configuration parameter set....................................................................................... 113
8.3 ROBIN statistic.......................................................................................................................114
8.4 ATLAS detector parameters....................................................................................................116
8.5 Glossary..................................................................................................................................119
List of figures..................................................................................................................................... 121
List of tables....................................................................................................................................... 122
Bibliography.......................................................................................................................................123
Acknowledgements............................................................................................................................ 129
6
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 1 - Introduction
1
Introduction
The research objective of Particle Physics is the structure of matter. According to the Standard Model
[SMS1][SMS2][DISSFLICK] there are 12 particles which form matter and 4 force carrier particles
which describe the interaction between the matter particles. Matter particles are distinguished in
quarks and leptons and further into 3 generations. Each generation is composed from a pair of each
type of particles. Only the 4 particles of the first generation – the electron with its neutrino and the up
and down quarks – build the matter we are aware of. The interactions between particles are called
forces and are bound to the the exchange of force coupling particles, the bosons. Two of the forces –
gravity and electromagnetism – are quite commonly known. The two others – the weak and the strong
force – are related to interactions between and inside nuclear particles, like protons, neutrons and
quarks. The strong, weak and electromagnetic force cover a relative strength range of about 1:100, but
the strength of the gravity is 43 orders of magnitude smaller than the strength of the strong force. This
is the reason why gravity is ignored1 by many of the Standard Model calculations. We have
experimental evidence for the existence of the bosons carrying the first three forces. Differently
however for gravity, the boson carrying the mass – the Higgs-boson – is still undiscovered.
To find the Higgs and to explore other important issues of particle physics – like SUSY [SUSY], CP
violations [CPV] or QCD [QCD] – a huge machinery has been developed: the LHC [LHCJINST] at
CERN. At the LHC two proton beams are accelerated and are travelling in opposite direction in the
LHC ring. At 4 experimental stations, ATLAS [ATLJINST], ALICE [ALICEJINST], CMS
[CMSJINST] and LHCb [LHCBJINST], the beams can be focused such that the protons collide with a
maximum energy of 2 * 7TeV, an energy level close to that at the origin of the universe.
Figure 1: Experiment Scenario
Each of the experimental stations, where ATLAS is the largest of, employ very complex equipment –
called a detector2 – to find all the particles which are produced during and immediately after the
collisions. The ATLAS detector is a 44m long, 25m high, 7000t heavy precision measurement
instrument consisting of 7 different sub-detectors3, generating data on roughly 100 million channels at
1 This is clearly opposite to the human common sense, which easily accepts gravity as the strongest and most
obvious force, while it is not aware of the strong and weak forces.
2 The “detector” is in fact a collection of different detectors, sensitive for different particles like electrons and
muons.
3 The sub-detectors are: Muon chambers, semiconductor tracker, transition radiation tracker, pixel detector,
LAr electromagnetic calorimeters, LAr hadronic calorimeters and tile calorimeters.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
7
Chapter 1 - Introduction
a interaction rate of the protons of 1GHz4. The event recording rate however is limited to 200Hz at an
event size in the order of 1MB. To achieve the required data reduction of 5x106 a complex and
efficient trigger and data-acquisition system (TDAQ) is needed. A simplified view of this general
scenario is shown in Figure 1.
Traditionally, trigger and data-acquisition systems in high-energy physics (HEP) utilise custom
electronics in many areas, for example to interface to the detectors, to satisfy extreme performance or
density requirements or because the electronics is exposed to radiation. There are also several
disadvantages to custom electronics, in particular cost, long development times and a long-term
dependency on particular know-how and components. Considering the development time of 15 years,
the operation time of another 10 years or more and the cost, there was a strong desire to use COTS
wherever possible to build the ATLAS TDAQ system. The ATLAS solution to this dilemma in the
TDAQ area is to equip standard PCs (the COTS) with custom modules, which provide the missing
performance and functionality.
ATLAS TDAQ uses a layered trigger architecture to reduce the amount of data to be transferred. The
levels are called level-1 trigger (L1), level-2 trigger (L2) and event-filter 5 (EF), the latter two form the
higher-level triggers (HLT). The L1 is closely coupled to the front-end electronics of the sub-detectors
and uses custom electronics to get the event rate down to 100kHz. L2 and EF are based on farms of
commodity PCs. Unlike many HEP experiments the L2 of ATLAS does not operate on the full detector
data volume but reduces the amount of data with two mechanisms – the region-of-interest (RoI)
principle and the sequential selection strategy. Figure 2 shows the traditional approach on the left,
where data rate and volume are reduced layer by layer while data is pushed through the TDAQ system
and the ATLAS approach on the right, where the the data is stored after the first step (L1) in the ROS
and L2 and EF request only a fraction of the data.
Figure 2: TDAQ levels
The ROS, again a farm of PCs, implements the buffering capabilities of the DAQ and interfaces to the
detector/L1 system on one side and to the HLT system on the other side. The detector/L1 side
generates fragments of the event data on 1600 channels and transmits them at the L1-rate of up to
4 The proton beams are structured in bunches and the bunch crossing frequency is 40MHz. For individual
protons the collision rate becomes 1 GHz.
5 The third trigger level (L3) is called Event Filter (EF) in ATLAS terminology.
8
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 1 - Introduction
100kHz through unidirectional6 optical links (ROL) with a nominal bandwidth of 160MB/s. L2 and EF
subsequently request portions of the data from the ROS. Typically, L2 requests occur at a higher rate
but address only a fraction of the fragments while EF requests occur at a lower rate but address full
events. Rate and bandwidth exceed the capabilities of commodity PCs by far.
To achieve the required performance, a custom FPGA-based I/O-coprocessor – the ROBIN – was
developed. Early studies [DJFROB] showed, that the handling of a single ROL using a combination of
re-configurable logic and processor is possible7. Subsequent technology studies led to the development
of different single channel modules, followed by a dual-channel ROBIN-prototype. The final
implementation of the ROBIN concentrates three input channels on a 64-bit PCI card. The architecture
and hardware design of the ROBIN is based to a large extent on systems developed previously by the
author (MPRACE-1, see 8.1.1 ) or to which he has contributed significantly (µEnable, see 8.1.2 ).
In addition to the baseline readout scenario via the PCI bus it also supports the switch-based readout
scheme via an integrated 1G-Ethernet port. Due to the high performance requirements a custom datapath controller is needed, which is implemented in a 2M-gates FPGA. FPGA technology enables to
create efficient and optimised I/O interfaces, memory controllers and logic functions in a single
device, while keeping flexibility and maintainability. An attached processor implements the
management, messaging and monitoring functionality.
This work describes the development, implementation, production and installation of the ATLAS
ROBIN. The results show that the ROBIN as a component satisfies the requirements with respect to
performance and cost. In addition, the ATLAS ROS, which is the framework for the ROBIN, has
demonstrated to reach its performance goals. The low failure rate observed during approximately one
year of operation prior to the LHC start-up in September 2008 demonstrates the reliability of the
ROBIN.
This thesis is structured as follows:
•
Chapter 2 provides an overview of typical HEP data-acquisition systems, with a focus on the
CMS experiment.
•
The ATLAS Trigger and DAQ system is described in chapter 3 .
•
Important technological aspects – in particular the FPGA technology – used for the
implementation of the ROBIN are introduced in chapter 4 .
•
The details of the ROBIN: development, implementation, production and commissioning are
described in chapter 5 .
•
The results from stand-alone and system tests are presented in chapter 6 .
•
Chapter 7 concludes with an assessment of the achieved goals and an outlook for the
upcoming ATLAS operation and upgrade phases.
•
The appendix provides information on previous systems and some additional tables relevant
for the operation of the ROBIN.
6 The ROLs are physically bi-directional, but the sole use-case for the return path is flow-control.
7 The early prototypes didn't have any of the operational monitoring features present on the current ROBINs.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
9
Chapter 1 - Introduction
10
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 2 - Data Acquisition Systems
2
Data Acquisition Systems
This chapter presents the main issues typical data acquisition system have to cope with. The “PUSH”
data transfer model followed by ALICE, CMS and LHCb is introduced and its distinction from the
“PULL” model employed by ATLAS. The implementation of CMS – the largest partner experiment to
ATLAS – is shown in detail and the characteristic parameters of ALICE and LHCb are presented.
In short, the task of a data-acquisition system is to collect data from an arrangement of sources and to
deliver it to a unit which performs storage and typically analysis. In the case of ATLAS and the other
LHC experiments the individual DAQ systems have to deal with some or all of the following
challenges:
•
Large number of input channels
•
Large data volume
•
High rate of events
•
High connectivity and throughput requirements on networks
•
High performance requirements for event processing
•
Buffering required to compensate latencies
•
Low cost, longevity, scalability, etc.
2.1 LHC Experiments
The LHC proton-proton collider is situated at CERN, Switzerland in a 27km long circular,
underground tunnel (see Figure 3). There are four main experimental sites: ATLAS, ALICE, CMS and
LHCb. ATLAS and CMS are general purpose experiments for the study of proton-proton collisions.
ALICE is specialised for the study of heavy-ion collision, LHCb is specialised for the study CP
violations in B-meson decays. ATLAS and CMS are the two largest experiments, which pose similarly
high requirements on the Trigger/DAQ system: L1-rate of 100kHz, event size of 1MB and a storage
rate of 100Hz, in other words an input bandwidth of 100GB/s and an output bandwidth of 100MB/s.
ALICE and LHCb have different demands – higher L1-rate at smaller event size or vice versa. The
DAQ of LHCb runs at an event rate of 1MHz and an output rate of 2kHz, the average event size is
35kB [LHCBDAQ]. In contrast, ALICE DAQ has a low input rate of around 8kHz but event sizes can
be as large as 85MB. ALICE operates in different modes, with and without trigger, the maximum
sustained output bandwidth is assumed to be 1.25GB/s.
The experiments ALICE, CMS and LHCb follow the traditional “PUSH” model for data transportation
in the DAQ system, although different options have been looked at during the development phases
(e.g. so called “phased” readout in LHCb [LHCBTB]). The main characteristic here is the fact, that for
every event the entire data volume is unconditionally transported – typically over several layers of
different media – to a node which assembles the full event prior to the recording on permanent storage,
which is based on a trigger decision. This leads to very high bandwidth requirements on the
interconnect infrastructure. Also, low latency transmission is needed to avoid substantial intermediate
buffering. In ATLAS and CMS for example the stages feeding the DAQ dataflow system provide
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
11
Chapter 2 - Data Acquisition Systems
virtually no buffering beyond the 3 to 5µs latency of the L1 trigger. Many networking technologies
have been investigated during the development phases of the LHC experiments with an emphasis on
Ethernet due to its dominance in the mass market. To overcome the limitations of its technical
properties – in particular the missing reliability, which is frequently compensated by reliable protocols
like TCP/IP at the expense of long latencies – optimisations had to be implemented by the two
experiments. CMS uses a reliable initial network layer to form sub-events which are then transported
via a reliable protocol over GE. ATLAS reduces the required network load with the help of the
additional L2 and ROS stages, which allows to tolerate an unreliable network. For ALICE and LHCb
the total network bandwidth is lower hence better adapted to GE technology. The total connectivity of
the dataflow networks is in the order of 1500 ports for ATLAS, CMS and LHCb and 350 ports for
ALICE.
Figure 3: LHC at CERN*
The “PUSH” model was considered by ATLAS initially as well, but then the experiment has made a
different choice by implementing a true “PULL” model, in order to take full advantage from the
sequential selection scheme and the RoI principle (see section 3.1 ). Here, the detector data are
captured directly after L1 in an intermediate buffer in the ROS which forwards them only upon request
to the HLT and DAQ system. This approach reduces the amount of data to be transferred to about 6%
of the initial data volume after L1 for the typical case. This reduction in the total throughput of the
dataflow enables the use of an all-commodity network infrastructure. The drawback is the increased
12
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 2 - Data Acquisition Systems
system complexity due to the additional flow of control messages (data request and delete messages),
the additional intermediate buffering stage and L2 trigger.
All LHC experiments employ a custom component to receive data from the sub-detectors over
dedicated links and to feed them – directly or via a PC – into a network infrastructure. Due to the
transport model used in ATLAS this component – the ROBIN – has a complexity far higher than that
of the other experiments. Technically, the ATLAS ROS could be tuned from its regular operation mode
(reducing the data volume by a factor of 10 roughly) up to a full readout mode (forwarding the entire
data volume), of course at the expense of additional components. Also, conversion to “PUSH” mode
would be viable (however is not foreseen) by re-programming of the existing components.
The “PULL” mode is not viable for the DAQ systems of the other LHC elements, as the network
infrastructure is not prepared to handle the required rate of request messages and there is no
intermediate buffering capacity available to cover the latencies of the HLT systems. Investigations
during the upcoming SLHC upgrade programme will certainly revisit both models for all experiments.
2.1.1
CMS
The CMS experiment is a large multi-sub-detector precision measuring device for charged particles
and photons. The momentum of charged particles is measured via the curvature of particle tracks,
caused by the bending force of a strong, superconducting magnet. The inner tracking detectors (pixel,
silicon strip) and the hadronic and electromagnetic calorimeters are placed inside the 13m long
magnet. Four muon-detecting stations are located externally to the magnet. The arrangement is shown
in Figure 4. When operating at design luminosity approximately 1.000 particle tracks per bunchcrossing (25ns) will pass the inner tracking detectors. This requires a high resolution of the pixel and
silicon strip sensors, leading to high power and hence cooling requirements. However, the resulting
installations (cables, cooling pipes) introduces unwanted effects8, so the design is a compromise.
2.1.1.1 Trigger
CMS uses a 2-level trigger system. The L1 is implemented in custom hardware and reduces the
40MHz bunch-crossing rate to an L1-event rate of 100kHz. The generated data volume is in the order
of 100GB/s, distributed over 626 data sources from the various sub-detectors. This entire data volume
has to be transferred to a large computing farm running the HLT algorithms [CMSTRIG].
Subsequently, filtered events are sent at 100Hz to permanent storage.
8 All material in the detector can cause multiple scattering, bremsstrahlung, photon conversion and nuclear
interactions.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
13
Chapter 2 - Data Acquisition Systems
Figure 4: CMS Detector [CMSJINST]
2.1.1.2 DAQ
The DAQ architecture responsible for the transport of data is shown in Figure 5. All detector channels
– approximately 55 million – are assembled into 626 event fragments of 2kB size in the detector frontend electronics and buffered during the L1 latency of 3.2µs in the front-end-drivers (FEDs). The event
fragments are pushed at L1-rate from the FEDs through LVDS cables9 into 458 front-end read-out
links10 (FRLs). A Myrinet11 based FED-builder (FB) network creates 72 super-fragment [CMSSFB]
streams – each composed of 8 fragments from different FRLs – which are directed to 72 readout units
(RUs). Full events are then built by combing the data from all 72 RUs via the readout-builder-network
in the builder-units (BUs), which also run the HLT algorithms via the filter-units (FUs) applications.
As a pure “push” architecture CMS DAQ does not require any significant buffering of the event
fragments but a very large network bandwidth.
This architecture has an inherent scalability provided by the 8x8 switches of the FB layer. Every FB
switch can route the fragments from the 8 input links to up to 8 output links, for example based on the
L1ID. CMS uses this feature to build a “sliced” DAQ system, where the total performance can be
tuned via the number of slices attached to the FB layer (see Figure 5:side view). A single slice consists
of 72 RUs, a builder-network switch and 288 BUs/FUs and is able to process 12.5kHz of L1-rate. The
9 The LVDS links follow the S-Link64 specification [SLINK64]
10 The canonical number for FRLs is 576. Only FRLs are equipped with 2 input links which connect to FEDs
with lower output bandwidth, such that the nominal input bandwidth of 200MB/s per FRL is not exceeded.
11 http://www.myri.com/open-specs/index.html
14
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 2 - Data Acquisition Systems
full performance of 100kHz L1-rate is obtained with 8 slices, which corresponds to 576 RUs and 2304
BUs.
Figure 5: CMS DAQ [CMSJINST]
2.1.1.3 Readout Builder
The second stage of the CMS event-builder – RUs, builder network and BUs – is made from COTS
components: PCs and a Gigabit-Ethernet (GE) network. Due to the layered and sliced architecture the
required connectivity is only 72 x 288 per readout-builder network (instead of 512 x 2304 for a full
single layer event-builder). However, GE does not match the input bandwidth (200MB/s) on the RUs
from the Myrinet FED builder, hence a multi-rail implementation [CMSGBE] with up to 4 GE-NICs
per node was chosen, leading to an aggregate bandwidth of 500MB/s. To achieve a high network
performance the Ethernet packet size had to be increased to 7kB from the standard MTU size of 1500
byte. Results from a 16RUs x 60BUs test-bed12 using 2 GE rails show a throughput for an RU well
above the required 200MB/s for the standard super-fragment packet size of 16kB. The steering of the
super-fragments into the RUs and from the RUs to the BUs is controlled by an event-manager (EVM)
which also checks the memory allocation at the RUs and eventually requests to reduce the L1-rate via
an interface to the global trigger processor (GTP). The BUs send readout-request to the EVM which in
turn provides event-tag information to the RUs, which then send the super-fragments to the indicated
BUs.
2.1.1.4 FED Builder
The first stage of the CMS event-builder is composed of the FRLs and the Myrinet FB. Myrinet was
selected due to the inherently reliable transmission which is achieved by using a physical media13 with
12 The test-bed used dual single-core XEONs with 2.6GHz.
13 Initially, Myrinet used parallel LVDS transmission. The current implementation is based on optical
transmission with 2Gbit/s.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
15
Chapter 2 - Data Acquisition Systems
a low error rate, built-in CRC error detection and an efficient flow-control mechanism using a “slac”14buffer. Myrinet packets can be of any length. In addition, all Myrinet network adapters and switches
are equipped with a user programmable RISC processor which is used by CMS to add custom
functionality. The FB connects all FRL sources (up to 512) with all 572 RU destinations via 8x8 port
switches15. The CMS routing strategy implemented on each of the 8x8 switches combines one
fragment per input into a super-fragment. Super fragments are sent to the RUs attached to the output
ports depending on the event-number and a pre-loaded routing table, which in turn depends on the
number of DAQ slices (every output represents a slices). As the FRLs on the input are equipped with
dual-link Myrinet NICs the FB is composed from two independent “rails”, each connected to one of
the FRL outputs. The two-rail structure is shown in Figure 6.
Figure 6: CMS FED-Builder Rails [CMSJINST]
Considering its location at the interface to the detector the FRL is the CMS component which is most
equivalent to the ATLAS ROBIN – although it does not provide significant buffering. Furthermore, it
has certain technical similarities, namely the combination of FPGA, embedded processor and host
CPU. The FRL [CMSFRL] is a single width cPCI16 card which consists of a custom base module,
attached to the cPCI bus plus a commercial dual-port Myrinet network interface adapter (NIC17),
plugged on to the base module via a standard PCI connector (Figure 7). The main FRL FPGA18 merges
the fragments from the two S-Link64 inputs (with the help of a small amount of external memory),
checks the CRC and writes the combined data in blocks of fixed length to the Myrinet NIC.
Additionally, the FRL provides monitoring features like fragment size histograms and can take data
samples. These auxiliary functions are available via the cPCI bus to the crate controller CPU which
14 The Myrinet slac buffer is basically a FIFO with low and high watermark indicators, which are used to stop
and start incoming transfers.
15 In practice a larger switch is used to implement multiple independent 8x8 groups.
16 Compact PCI (cPCI) is the industrial version of PCI, see http://www.picmg.org/test/compci.htm
17 http://www.myri.com/vlsi/LanaiX.Rev1.1.pdf
18 The main FPGA is an ALTERA EP1S10 device with 10k logic elements (LEs), 60 block memories and 6
DSP blocks.
16
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 2 - Data Acquisition Systems
controls up to 16 FRLs per crate. The nominal bandwidth on the input and output is 200MB/s, which
is easily met by the S-Link64 (400MB/s). The Myrinet NIC is a dual-link version and both links
together provide 4Gbit/s.
Figure 7: FRL image and block diagram[CMSFRL]
Although the NIC is a commercial device, it is used together with a custom firmware which allows it
to communicate with the main FPGA on the base module via a private PCI bus. The FPGA deposits
the merged event fragments into the local buffer of the NIC and instructs it to transmit the packets by
writing a DMA descriptor. At the system level, a throughput of 300MB/s per FRL for varying size
fragments with 2kB average has been measured [CMSSFB], well above the requirement of 200MB/s.
2.1.2
ALICE
The ALICE dataflow architecture is displayed in Figure 8. Raw data are collected close to the detector
in front-end readout cards (FERO) and transmitted via optical links (DDL) to data receiver cards (DRORC), two of which are hosted each by one local data concentrator (LDC). The LDCs build sub-
Figure 8: ALICE dataflow [ALICEJINST]
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
17
Chapter 2 - Data Acquisition Systems
events (similar to CMS super-fragments) and output them to a GE network. Full events are assembled
by global data concentrators (GDC) and ultimately shipped to mass storage, eventually based upon
decisions from the HLT farm. The HLT trigger is fed from a subset of the detector only, via bypass
links each established by one of the D-RORC DDL channels.
The ALICE component which is most equivalent to the ROBIN is the D-RORC [DRORC], which
receives two optical inputs at 200MB/s each and stores the data into the memory of the host PC.
Alternatively, one of the links can be used to send a copy of the input data to the HLT farm. As can be
seen from the block diagram [ALICETDR] in Figure 9 the D-RORC does not contain any local
memory or processor, but is just an I/O extension to the PC. Measured throughput into host-memory
reached 484MB/s, well above the combined bandwidth of two optical links.
Figure 9: ALICE D-RORC [DRORC]
2.2 LHCb
The dataflow architecture [LHCBADD] of LHCb as shown in Figure 10 is relatively simple, as it does
not use a separate L119 concentration layer. Instead, each of the 310 front-end (FE) units [LHCBTELL]
Figure 10: LHCb dataflow[LHCBADD]
19 In LHCb the first trigger level is called L0. However to avoid confusion the term L1 is used in this document.
18
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 2 - Data Acquisition Systems
located close to the detector collects data from a combination of up to 64 analog or 24 digital20
channels respectively and transmits L1 accepted events via 4 built-in GE output ports to the readout
network, which provides a total connectivity in the order of 2500 ports maximum. LHCb aims to keep
the load on the GE network below approximately 70%, as experimental studies have shown that the
risk for message loss is close to zero for loads below this value.
The FE card – called TELL1 – is a 9U VME card (Figure 11, right) which features several different
I/O interface in addition to the data input and output ports. The ECS ports connects to the slow-control
system of the experiment, the TTC port receives the central timing and control information. A
throttling mechanisms slows down the L1 accept rate once the local buffers, which need to cover the
4µs L1 latency, become saturated. The PP-FPGAs (Figure 11, left) perform sub-detector specific
processing of the raw detector data, while the SyncLink-FPGA is responsible to assemble and transmit
event fragments.
Figure 11: LHCb TELL1 [LHCBTELL]
Compared to the other experiments the TELL1 cards looks rather like the combination of front-end
readout and DAQ interface, an approach which was also investigated by ATLAS in the course of the
ROB-on-ROD project (see chapter 3.3.4 ).
2.3 Summary
The large LHC experiments CMS and ATLAS generate a few 100 particles every 25ns when operating
at design luminosity. Particle tracks are recorded by different detector subsystems via 50 to 100
million sensor channels. A custom first level trigger located close to the detectors is used to select
interesting events and provide them to the DAQ/HLT subsystem, which in turn is characterised by a an
input rate of 100kHz and an input bandwidth in the order of 100GB/s. The event building processes
require to connect roughly 1.000 sources to 1.000 destinations. The latter filter the events according to
trigger algorithms, leading to an overall reduction in bandwidth and rate towards mass storage by a
20 The nominal speed of the digital inputs is 1.25Mb/s.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
19
Chapter 2 - Data Acquisition Systems
factor of 1000. All LHC DAQ systems are in general based on commodity technology - GE networks
and standard PCs - but need some custom components at the input stage. In particular for CMS and
ATLAS a scalable system was required due to financial issues, which allows to operate the system
with reduced resources at reduced performance.
The CMS approach as described in this chapter transports the entire data volume – nominally events of
1MB size at a rate of 100kHz – via a layered architecture, following the traditional "PUSH" model.
The initial layer employs 458 custom FPGA-based data converters (FRLs) and a Myrinet builder
network (FB) made from 72 8x8 switches. The FB provides data from all sources on 8 parallel outputs
which feed up to 8 slices of the subsequent stage. Each of the slices consists of 72 readout-units (RU)
receiving Myrinet packets at 200MB/s, a GE network and 288 builder units (BU), which assemble the
full events and run the event filter algorithms. The performance is scalable at 12.5kHz per slice.
ALICE and LHCb have somewhat different characteristics in terms of rates and event sizes and lower
requirements with respect to total bandwidth. However, both share the same “PUSH” model for the
dataflow as CMS does and employ similar architectures and technologies for the custom DAQ
components.
In contrast, ATLAS introduced the intermediate L2 layer operating on RoIs – typically only a few kB
at a rate of 100kHz – and transporting full events – about 1MB each – at a low rate of a few kHz. This
architectural decision was taken with the aim for a full commodity dataflow network and under the
assumption that processing power at the L2 will be relatively inexpensive at the time of operation.
All four experiments use custom FPGA components at the boundary between the sub-detectors and the
dataflow system, which translate from proprietary signals into some networking standard. In case of
ALICE and CMS the functionality is basically limited to this translation step. The LHCb component
integrates additional functionality related to data processing and first level event selection. The
ATLAS dataflow architecture is unique in its demands for buffering at the input and communication
capabilities to serve the “PULL” model and requires high-performance FPGA technology combined
with high-performance embedded processing – which are realised by the ROBIN component. The
ATLAS approach is explained in the following chapter.
20
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
3
ATLAS DAQ
ATLAS is the other “large” experiment at the LHC, with requirements very similar to CMS (see
section 2.1.1). This chapter introduces the ATLAS detector with the focus on the architecture of the
ATLAS TDAQ system. The unique ATLAS approach to reduce the throughput requirements by a
significant amount using the RoI-principle and sequential-selection in the L2 trigger stage is
explained, due to the impact on the architecture of the dataflow. The baseline dataflow system of
ATLAS and the ROS are described in detail, which define the requirements on and the environment of
the ROBIN.
3.1 System View
The ATLAS detector is – like CMS – composed from different sub-detectors to identify charged
particles (muons, electrons, …) and to measure other observables (vertex displacement, missing
transverse energy, …). The layout of the magnets is different from CMS, which uses a single large
solenoid magnet while ATLAS has a an inner solenoid and a group of toroid magnets further away
from the centre. A sketch of the 7000t heavy detector is shown in Figure 12.
The seven sub-detectors of ATLAS together generate data on roughly 100.000.000 channels, the vast
majority of belonging to the pixel detector. These channels are collected by electronics on or close to
the detector and combined into 1600 ROD modules. The RODs are also connected to the L1 and the
timing system of ATLAS and format data plus event identification into ROD fragments. A simple
Figure 12: ATLAS detector*
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
21
Chapter 3 - ATLAS DAQ
unidirectional link, the S-Link [SLINK], is used to connect the RODs with the TDAQ system. From
the TDAQ point of view the ATLAS detector is a data source with 1600 channels, 100GB/s
bandwidth21 and 100kHz event rate. The task of ATLAS TDAQ is to select and store only 0.1% of the
generated data volume.
The initial concept to implement the ATLAS TDAQ system as documented in the “ATLAS Technical
Proposal” [ATLASTP] is shown in Figure 13. After L1 a digital buffer memory is used to store L1
Figure 13: ATLAS TDAQ TP-version [ATLASTP]
data. An RoI-collection (RoIC) subsystem22 copies the RoI-portions of some sub-detectors to the L2
system. An RoI is derived from the geographical information 23 attached to the L1 event identifier
(L1ID), and defines a subset of a sub-detector, typically around 2%. The L2 is a modular system and
operates in two steps. In an initial feature-extraction step all sub-detectors are individually and in
21 The maximum bandwidth is 1600 channels * 160MB/s = 256GB/s. However, the nominal fragment size is
1kB and even less for quite some of the sub-detectors.
22 The digital buffer memories in Figure 13 have two output paths, the main towards “Readout/Event Building”
and a second one, which builds the RoIC, towards LVL2.
23 The particles generated in the collisions produce electrical signals in the detectors while they escape from the
the interaction point. Any signal generated on one of the detect channels is called a hit.
22
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
parallel analysed for interesting data. The features are subsequently combined by a global decision
step into the L2 decision, which is then distributed to the event building stage. At that time different
approaches for the L2 implementation were discussed and investigated in a demonstrator programme
[ATLDEMPROG] – a global L2 farm, the use of a local sub-farm24 per sub-detector and the use of
custom data-driven L2 processors based on FPGAs either replacing the local sub-farms or in a hybrid
fashion together with sub-farms. The different L2 options also affected the design of the readout
subsystem, e.g. the latency of the data-driven L2 was much lower than that of the global and local
farm approaches. However, it required the fast distribution of the RoI-information to the readout
buffers and the RoIC subsystem. The network technologies proposed for the L2 farm interconnects
were ATM, SCI and Fibre-Channel, none of which has a significant market share today. Also, three
different networks were to be used to transport event data from the buffers to the L2, from the buffers
to the L3 and to transport L2 decisions.
A significant simplification for the TDAQ architecture was achieved by the introduction of the
sequential selection strategy [SEQSEL] for L2. Sequential selection takes advantage from the fact that
all important physics events require the presence of signals in the Muon detector and the calorimeters.
Thus looking for Muons first allows to reject ¾ of the events at the L2, as shown in Figure 14. A
subsequent check for electrons, photons and jets enables rejection of another 60% of events. The
sequential selection strategy reduced the requirements on the system throughput and allowed to merge
the separate, sub-detector specific L2-subfarms of the initial architecture into a single, uniform L2
farm. The required processor farm size for the sequential execution of the L2 was estimated by
modelling [PAPMOD] to be in the order of 200 machines. However, sequential selection also has a
drawback, which is the increased complexity of the dataflow architecture. The traditional “push”mode has to be replaced with a “pull”-mode dataflow, which needs relatively advanced interactions
between the subsystems (see chapter 3.2 ).
A test-bed [PILPRO] was setup to verify the performance of the associated dataflow system with
different network and readout buffer implementations25 [ROBCPLX]. The range of technologies
investigated is summarised in Table 4.
Technology
Options
FPGA
High-end, provides Medium, I/O plus Low-end, I/O only
core functionality auxiliary functions
Processor
Host only
Local DSP
Local 32-bit
Microcontroller
Bus
PCI (PMC
mezzanine)
PCI (standard
format)
VME
Network
ATM
SCI
Gigabit-Ethernet
Optical links
2.5Gbit/s single
fibre
Multiple fibre
15Gbit/s (Paroli)
High-end SMP
host
Table 1: Pilot Project Technologies
24 At the time of the TP approximately 300 processors would have been needed for a local sub-farm.
25 During the pilot project phase the terminology “ROB Complex” was used which corresponds to the current
ROS.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
23
Chapter 3 - ATLAS DAQ
The pilot project defined the ROS to be the aggregation of a number of input units receiving data from
the RODs, associated buffers, some intelligent controller(s) to manage the event fragments and handle
requests and a number of output units interfacing to the L2 and EB network. The range of
configurations investigated covered the most simple one (single input unit, single controller, single
output) up to the “Active ROB Complex” [AROBC] where many input units were handled by a
powerful SMP host. The latter concept also included data-processing at the level of the readout
subsystem.
The results26 obtained during the pilot project phase showed that the requirements of the ATLAS
dataflow could be satisfied in principle with the proposed ROS architecture, however a solution was
needed in order to achieve the goals for density and cost. From the prototype implementations the ones
with a large fraction of functionality implemented in FPGAs provided the best performance, while the
processor-oriented suffered from the high rate. Two other areas were identified as potential
bottlenecks: the memory subsystem – in particular if shared by processor and buffer – and the parallel
bus, specifically the drop in available bandwidth for high message rates. All issues could be addressed
by the design of the subsequent ROBIN.
Figure 14: Sequential Selection [SEQSEL]
As a consequence, the ROS consisting of a standard COTS PC equipped with a number of FPGAbased ROBINs became the baseline implementation (see chapter 3.3.3) as documented in the ATLAS
Trigger and DAQ Technical Design Report (TDR) [ATLASTDR], which addressed the issues
mentioned above in the following way:
•
The use of COTS PCs reduces cost
•
Concentration of multiple input links per PCI card allows to build a compact system
26 The pilot project considered two operating conditions – low and high luminosity – of the LHC, with
corresponding L1-rates of 40kHz and 75kHz respectively. This is in derivation of the standard 100kHz/high
luminosity case assumed elsewhere in this work.
24
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
•
A central FPGA per ROBIN enables high-rate and high-bandwidth I/O handling
•
An auxiliary local processor provides flexibility
•
An auxiliary local network interface provides an upgrade path for bandwidth- or rate-intensive
request schemes
Figure 15: TDR ROBIN Blockdiagram
Figure 15 illustrates the TDR ROBIN design, with a central FPGA comprising the main data path from
the two input links to the two buffer memories and onwards to the alternative output interfaces PCI
and GE. FPGA technology – which is described in chapter 4 – enables to implement I/O interfaces,
memory controllers, direct-memory-access (DMA) engines and various other functions on a single
device in an optimised and highly efficient way while maintaining flexibility and upgradability. The
local processor is attached to the FPGA as well, but separately from the main data path. The final
design of the ROBIN is described in chapter 5.3 .
3.2 Dataflow
The ATLAS Dataflow system is responsible to move the detector data from the interface to the
detectors – implemented by the ROLs – up to permanent storage, attached via a standard network. A
set of subsystems constitute the main data path (Figure 16, center) from the RODs to the mass storage.
Additional subsystems provide control and configuration functionality (Figure 16, left and right). Due
to the specific properties of ATLAS TDAQ the dataflow uses a mixture of “push” and “pull”
mechanisms in order to transport data from one stage to the next one.
Starting at the RODs, event fragments are pushed to the ROBs. Concurrently, RoI information is
pushed from L1 to the RoIB and onwards via the L2SVs to the L2PUs. The L2PUs use the L2 network
to pull event fragments according to the RoI information via the ROS from the ROBs. The results
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
25
Chapter 3 - ATLAS DAQ
from L2 are pushed via the L2SVs to the DFM, which pushes “accept” decision to the SFIs and
“reject” decisions to the ROS. The SFI again pull all event fragments from the ROS via the EB
network, build the full events and push them to the EFPs. Finally, the events accepted by EF are
pushed via the SFOs to mass storage.
Figure 16: Schematic layout of ATLAS TDAQ system [ATLASTDR]
In the baseline dataflow model the ROBs are installed in ROS-PCs, each of which typically houses 4
ROBINs each representing 3 ROBs, so one ROS-PC serves 12 channels. The ROS-PC forwards all
relevant requests to the ROBs via the PCI-bus and combines the responses to SFI-requests into larger
fragments. In addition, the ROS-PC is responsible for configuration and monitoring of the installed
ROBs.
Alternative scenarios bypass the ROS-PC for L2 and/or SFI requests and pull fragments directly from
the ROBs, via their private Ethernet ports.
Apart from the bandwidth requirements ATLAS dataflow requires a network with good real-time
performance, high reliability and high rate capabilities in particular on the L2 network. The
requirements are largely comparable to the requirements of enterprise-size commercial networks
where Ethernet is the de-facto networking standard. However, there is one difference: ATLAS DAQ
requires low latency and reliability at the same time, while most of the typical Ethernet areas require
only one of them. Ethernet networks follow the best-effort principle and are not reliable per-se, like
26
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
Myrinet is. Client-server applications for example introduce reliability by using a reliable protocol like
TCP/IP over Ethernet, which can add significant delays to the transmissions. On the other hand,
multimedia applications like IPTV or video conferencing are sensitive to latency but quite tolerant to
packet loss and the unreliable UDP protocol is used frequently here.
The L2 system has an intrinsic latency of a few ms, caused by the execution time of the L2 algorithms,
which defines the buffering requirements at the ROS. Any additional latencies introduced by the
network increase the required amount of buffering capacity and make the system less stable. The use
of a reliable protocol like TCP is not a general solution, due to the relatively long retransmission
delays27 in case of packet loss and its unavailability 28 on the ROBINs. An additional requirement on
the reliability of the network at the low level is the use of multicast messages to distribute the DFM
results to the ROS. Although Ethernet does not appear to be the ideal candidate for the ATLAS
dataflow network from a technical point of view it was selected for reasons of cost, general
availability, ease of use and expected longevity, which is important for an operation time of 10 years or
more.
The total size of the ATLAS TDAQ system from the dataflow perspective is defined by the number of
external data sources and by the number of networking nodes. The number of ROBINs and ROSes
corresponds to the number of detector links and is fixed. The number of L2PUs and EFPs define the
trigger capabilities and thus influence29 the maximum L1-rate the system can handle. The expected
numbers for the individual components is given in Table 2.
Component
Instances in final system
Comment
ROL
1600
RoIB
1
RoI: 2% of full event
ROBIN
600
100kHz L1-rate
ROS
150
Separate network ports for L2 and EB
L2SV
10
L2PU
500
DFM
35
SFI
100
Full event size ~1.5MB
EFP
1600
~1 event/s per CPU
SFO
30
Dual-CPU systems, ~100 events/s per CPU
L2 network nodes 700
100 kHz, RoIs
EB network nodes 300
3 kHz, full events
Table 2: Data-flow components (TDR)
The large number of network nodes cannot be attached to a monolithic magic box which provides the
full connectivity for more than 1000 ports. Instead, the Ethernet approach for interconnecting nodes is
27 The TCP retransmission timeout is typically set to 3s, which results in delays in the order of seconds.
28 TCP requires to monitor every logical connection, which is too resource consuming on the ROBIN, where
several hundred simultaneous connections may exist.
29 Good linear scaling of L2, EB and EF documented in [ATLASTDR]
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
27
Chapter 3 - ATLAS DAQ
to cascade a number of units called switches, each serving up to a few hundred ports in a cross-bar
fashion. Significant effort has been put by ATLAS into the analysis of Ethernet behaviour and
equipment, well documented in [STANCU]. As mentioned above, the main issues with Ethernet –
apart from the connectivity – are latency and message loss. Both factors strongly depend on the
technology of the switches. In principle, a switch has a unit handling the input (ingress), a unit
handling the output (egress) and the cross-bar30 connecting input and output. The routing path over the
cross-bar is determined by the ingress unit from the header of every individual Ethernet packet, which
starts with a source and destination address identification. The switch constantly monitors all source
addresses and builds a map of addresses and ports. If a destination address is already in the table, the
packet is routed to the corresponding port, otherwise it is replicated and sent to all ports. The latency
introduced by this routing process is relatively small, as the evaluation of the routing path starts while
the packet is still being received from the source. An obvious problem is the case where multiple
source are sending to the same destination (so called funnel traffic) and such exceeding the egress
bandwidth limit. While Ethernet allows the switch to simply drop packets in this case the most
common solution is to queue packets at the input. If an input queue become full an Ethernet flowcontrol message is sent to the source, asking to pause the transmission. A common complication in
Ethernet switches is head-of-line blocking, which occurs when an congested egress is blocking an
ingress queue which contains also packets for another egress port. That egress can be idle in the worst
case despite the fact that packets are available for it in an ingress queue. Some switches improve the
situation by providing separate ingress queues31 for some or all egress ports but at high load both loss
and latency of Ethernet switch are inherently indeterministic.
Various test-beds have been set up to study the behaviour of the individual components and the
performance using different networking protocols. To achieve reasonable sizes for the test-beds
typically several components had to be emulated. For example, every ROBIN has built-in datagenerators able to provide event fragments of arbitrary size at full speed. FPGA-based data-generators
[GETB] [ETHERT] providing up to 128 ports and programmable NICs providing up to 16 ports were
used to create network traffic with relatively simple patterns at high rate. More complex traffic pattern
were created using PCs with special test programs. Using such systems, throughput in the order of
10% of the final system has been demonstrated for the baseline bus-based [BASEDF] readout
architecture. Large-scale tests [LARSC] were performed on computing clusters with several hundred
of machines, emulating different portions of the dataflow system up to the full size in most areas. The
analysis of the congestion management of various switches indicates that at loads up to 60% of the
nominal switch capacity the rate of lost message is virtually zero32 for random traffic patterns, if the
components are properly selected [ETHER]. Hence the capacity of the ATLAS dataflow network is
tailored such that the load on the switches stays below this margin. However, care must be taken that
the actual traffic patterns do not derive too much from a random distribution, such overloading certain
switch ports. A potential problem is the event building step where an SFI needs to get data from a large
number of sources. If the SFI would issue all data requests simultaneously the response packets would
30 There are also switches which use a shared-memory instead of the cross-bar, however the required memory
bandwidth poses a limit on the throughput and number of ports of such implementations.
31 This is also called virtual output queuing.
32 This is in line with the results obtained by LHCb (section 2.2 ).
28
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
certainly introduce head-of-line blocking in the switch with the associated effects, increased latency
and packet loss and ultimately reduced system performance. For example, a message loss rate of
0.01% results in a performance reduction of 20% of the SFI [STANCU]. The implemented strategy at
the SFI therefore uses a random delay after every data request.
Another consideration for the design of the ATLAS dataflow network was the communication pattern
– certain nodes communicate with each other and others do not. For example, the ROS nodes need to
communicate will all L2PUs and SFIs but the SFIs never communicate with the L2PUs. Also, nodes of
the same kind do not communicate with each other. The analysis of the corresponding bandwidth
requirements shows that concentrating switches can be used in to aggregate ROSes or L2PUs. The
small concentrating switches are then connected to the central switches, which also connect to the
SFIs. The uplinks from the concentrators to the central switches use either several GE or a single
10GbE link. Figure 17 Shows the layout of the dataflow network using concentrator switches for the
L2PUs, for the control nodes (DFM, L2SV and pROS) and for some of the ROSes. Two large central
switches with a nominal capacity in the order of 250Gbit/s each build the core of the network. L2 and
EB traffic is mixed in this scenario which provides also a certain degree of fault tolerance, as the
system can continue to run even after the failure of a central switch, although at a lower rate. The
alternative scenario where one central switch is used for EB and the other one for L2 was preferred
earlier as it keeps the subsystems separate, however at the expense of missing flexibility and fault
tolerance.
Figure 17: Dataflow network [STANCU]
The networks have been characterised using UDP and TCP and both protocols can be used
concurrently in the system. In general, network latencies and message loss can be kept at an acceptable
low level using standard – sometimes selected – network components [DFNET][DFROS]. Message
loss on UDP based L2 data-requests is handled at the L2 application level, by re-requesting the data or
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
29
Chapter 3 - ATLAS DAQ
by generating a “forced-accept” decision. Message loss on multicast delete messages is handled at the
ROS by a garbage-collection mechanism (see chapter 3.3.2 ). The buffering requirements introduced
by the L2 trigger latency depends on the actual event (due to the sequential selection) and was
estimated most recently to below 100ms in a relatively large test setup consisting of 134 ROS, 32 SFI
and 130 HLT nodes, using 4.000 simulated events [L2PROC]. The distribution of the processing time
is shown in Figure 18.
Figure 18: L2 processing times [L2PROC]
3.3 ROS
The ROS implements the buffering capabilities of the DAQ and the interfaces to the detector/L1
system on one side and to the HLT system on the other side. Event fragments are arriving from the
detector/L1 system through the ROLs at the L1-rate of up to 100kHz and with a nominal bandwidth of
160MB/s per link. The actual fragment sizes depend on the type of sub-detector and vary typically
between 400 and 1200 byte ([ATLASTDR], see chapter 8.4 ). Fragments generated for detector
calibration purposes may be much larger (approx. 10 - 100kB), however occur at a very low rate. A
full event is composed from all fragments corresponding to the same L1ID and has a typical size
between 1MB and 1.5MB.
A baseline ROS architecture [Ibid.] has been developed, which concentrates a significant number of
ROL channels into a single PC via the PCI-bus, satisfying the typical requirements. For a 12-channel
ROS-PC the total L2 request rate varies between 300Hz for the TRT sub-detector and 18.6kHz for the
electromagnetic calorimeter [Ibid.]. While under typical conditions the L2 and EF requests average to
6kHz per channel at a L1-rate of 100kHz the theoretical worst-case rate 33 on an individual channel is
around 21kHz [MADROS]. The bandwidths related to the typical and worst-case request rates are in
the order of 10MB/s and 33MB/s per ROL respectively. Enhanced ROS performance is achieved
either by reducing the number of channels per PC or by employing additional networking connectivity
33 The 21kHz are composed of 18kHz L2-rate plus 3kHz EB-rate. Under normal conditions it is very unlikely
that the full L2-rate goes to a single channel.
30
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
bypassing the PCI-bus. The ROBIN – the only custom component of the ATLS dataflow apart from
the RoIB – was designed to handle and buffer the input data with the flexibility to interact with the
ROS-PC as well as directly with the dataflow system via a private network interface. In addition to the
tasks above, which are related to the transport of detector data, the ROS interfaces to the configuration
database and run control, the monitoring system and the detector control system (DCS) [ATLDCS].
The requirements on the ROS are summarised in a ROS user requirements document [ROSURD].
3.3.1
Event handling
In ATLAS, events are identified by a 32 bit event number composed from a 24 bit identifier generated
by the L1 trigger – the primary L1ID – plus an 8 bit event-counter-reset (ECR) value, incremented by
the RODs upon the wrapping of the event identifier to 0. For simplicity the event number is normally
and in this thesis referred to as L1ID. A design constraint limits the frequency of the ECR to the range
from 0.1Hz to 1.0Hz. As a result, the maximum time between two zero-crossings of the event number
is 256s, equivalent to 25.6 million events at a L1-rate of 100kHz. The minimum time is 25.6s. As a
typical ATLAS run (a period of continuous operation) can extend to several hours the L1ID is not
necessarily unique for all events and an additional mechanism needs to be put in place for event
identification. As the minimum time covers the range of L2 and EF latencies the additional
information is inserted at the event building stage by adding appropriate time-stamp information. At
the ROS level the limited amount of buffering space requires to delete events as soon as possible. This
is done by explicit delete messages distributed by the DFM, after the event has been rejected by the L2
or processed by the EF. To reduce the rate for delete messages they are sent out via a multicast
mechanism typically in groups of 100.
As stated in chapter 3.2 the dataflow system is designed to minimise packet loss on the network
switches, however losses are not fully prevented. While data requests are point-to-point interactions
and can be protected by a reliable protocol this is not the case for delete messages. Lost delete message
lead to orphaned fragments in the ROS and reduce the amount of buffer space available. While the
event numbers at the ROS level restart at 0 after a maximum of 256s it cannot be guaranteed that all
orphaned events will be replaced, as the L1ID does not have to be strictly sequential. Therefore, a
“garbage collection” mechanism is required to clean up the buffer. The implementation of this
mechanism requires to distribute the “oldest” valid L1ID in the dataflow system, which is piggybacked to the delete messages. The loss of a message is detected by a jump in the message sequence
numbers. Once a lost delete message is detected, the ROS compares the oldest valid L1ID to the most
recent L1ID received from the ROLs, creates a range of valid L1IDs and deletes all fragments with
event numbers outside of this range. The actual garbage collection procedure is executed on the
ROBINs, which need to do a time-consuming scan of their entire buffer in order to build the list of
stored fragments. To avoid excessive load the garbage collection is executed only when the buffers on
the ROBIN have reached a certain filling level.
3.3.2
Configuration and monitoring
Every ATLAS run is associated with a set of configuration parameters, stored in the global ATLAS
configuration database. The range of parameters is very broad and includes calibration values for
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
31
Chapter 3 - ATLAS DAQ
detector front-end electronics, RoI-mappings and IP-addresses. A number of these parameters controls
the behaviour of the ROBINs, distinguished into regular parameters and expert parameters. The
regular parameters include buffer memory page sizes34, truncation limit, IP-address, channel identifier
etc. The expert parameters include values which need to be modified in order to enable or disable
particular functionality required to perform specific tests. During a regular run all expert parameters
are set to their default values. The configuration of the ROBIN is controlled by the ROS-PC via a the
normal requests/response mechanism.
The ATLAS TDAQ system requires a detailed online view of all activities in the system, in order to
properly react to any malfunction. Thus, every subsystem has to provide functions related to
operational monitoring. At the ROS level, operational monitoring gathers statistics information of
received and lost messages, of buffer pages, processed fragments and of errors and histograms of
buffer occupancies and fragment sizes. Most of this information is prepared by the ROBIN and
transported to the ROS-PC via the regular requests/response mechanism.
3.3.3
Baseline bus-based ROS
According to the baseline bus-based35 architecture, the ROS is built from 150 PCs installed into racks
with up to 12 PCs each (see Figure 19, front and rear view of rack). Each ROS-PC36 attaches to 12
ROLs with a total input bandwidth of almost 2GB/s at a fragment rate of 1.2MHz. On the DAQ/HLT
side the ROS has to handle the requests from L2 and EF for event data 37 and event rejection38.
Connectivity to the L2 and EB networks is implemented with a 4-port NIC, which uses 1 port for each
of the networks in the default configuration.
An additional NIC port is used for the operating system's network interconnection and for control and
configuration. The bandwidths corresponding to the typical conditions are 60MB/s per network. The
performance requirements as documented in the ATLAS TDR relate to fractions of the L1 input rate
and translate for the standard ROS to 4kHz EB plus 12kHz L2 (RoI size of 1) at 100kHz L1 rate and
1kB fragments. A “hot-spot” condition was defined with 17% of L2 (RoI size of 2) and a fragment size
of 1.4kB. Early measurements for a typical 12 channel ROS, equipped with data emulators, showed
that the standard requirements were within reach (94kHz L1 rate achieved). For the “hot-spot” ROS
either the EB rate had to be lowered to 2% or the L1 rate to 75kHz. As the “hot-spot” condition applies
only to a few ROS-PCs attached to the electromagnetic calorimeter, the proposed solution at the time
of the TDR was to reduce the number of ROBINs to 3 or even 2 in the few ROS-PCs affected.
34 The buffer memory page size on the ROBIN has a default values of 2kB. Calibration runs for example can
use larger pages.
35 The baseline bus-based architecture uses the ROS-PC to interact with the DAQ/HLT system via the network
and to select and collect event fragments from the individual channels of the ROBINs. An alternative switchbased architecture allows the ROBINs to directly communicate with some of DAQ/HLT components via a
private network interface. This scenario is intended for special cases with more demanding performance
requirements.
36 The terminology of ATLAS TDAQ sometimes uses ROS synonymous for ROS-PC.
37 An L2 request addresses a single or a few channels, while an EB request addresses all channels.
38 Event rejection (delete) messages are issued by a separate DAQ component, the dataflow-manager (DFM).
32
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
Figure 19: Racks with ROS-PCs*
A further, even more demanding use case was defined after the TDR with a L2 request rate of 18 kHz
on all channels plus 3kHz of EB, at 100kHz L1 rate. Going even beyond that, a ROS could
theoretically be configured for 100% readout ratio, which would make it look somewhat like the CMS
FRL unit (see chapter 2.1.1.4 ).
3.3.3.1 ROS-PC
The typical ROS-PC comprises a single CPU, 1GB of main memory, a 4-port PCIe NIC and 4
ROBINs, as shown in Figure 20. The mainboard39 provides multiple PCI-X buses such that a
Figure 20: ROS-PC
39 Supermicro X6DHE-XB, http://supermicro.com/products/motherboard/Xeon800/E7520/X6DHE-XB.cfm
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
33
Chapter 3 - ATLAS DAQ
maximum of 2 ROBINs are connected to the same bus. This way, the maximum output bandwidth40 of
a ROBIN can be fully utilised. The NIC is placed on a separate PCIe bus. Both the motherboard and
the chassis were selected after testing a number of different machines for performance and stable
operation. Special attention has been paid to the quality of the power supply for PCI cards and to the
cooling facilities. Concerning the power supply it was observed that several motherboards did not
provide proper 3.3V if 4 ROBIN cards were installed. In some cases the voltage dropped below 3.0V,
which triggered the under-voltage reset circuitry of the ROBIN and prevented the boards from starting.
The thermal behaviour of the PC was tested by installing up to 5 custom “load-cards” (Figure 21) with
configurable dissipation between 1W and 28W. The air-flow generated by the front fan of the case
passes across all installed cards and exits through special openings at the rear of the case. The card
temperature has been measured for different cooling conditions (case open/fan on, case closed/fan on,
case closed/fan off) at a room temperature of 35°C. The temperature difference between the edges and
the centre of the cards on one hand and between cards at different positions is in the order of 10°C for
the situations where the fan is active, and the maximum temperature is around 65°C, which is
acceptable considering the high room temperature. If the fan is turned off, the maximum temperature
comes close to 80°C which is beyond the spec for many components. A regular monitoring of the
temperature of the ROBINs can be done with the on-board temperature sensor. Also, there are sensors
on the motherboard which can be used to detect failure of the cooling system in the PC.
Figure 21: PCI load board
3.3.3.2 Message passing
The ROS-PC runs a standard 32-bit Linux kernel, however with a patch41 that enables applications to
acquire a large amount of physically contiguous memory. The ROS application uses this memory to
build a pool of fixed-sized memory pages as destination buffers for ROBIN responses. The
communication between the main CPU and the ROBINs works in a messaging passing fashion: the
application sends a request to the ROBIN and the ROBIN returns a response. The requests follow a
standard format, which comprises the request code, a channel identifier, a sequence number, a
destination address, a length field and eventually any data related to the request. Requests are written
into a dual-ported memory area on the ROBIN, which is mapped into PCI memory space. A request
descriptor identifying the length and the location of the request is written to a separate memory area,
which is implemented by a FIFO on the ROBIN. The memory sizes of FIFO and dual-ported memory
40 The output bandwidth of a ROBIN is 256MB/s, the PCI-X bus supports 512MB/s at 66MHz.
41 The patch is called “bigphysarea”, see e.g. http://lkml.indiana.edu/hypermail/linux/kernel/0411.1/2076.html
34
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
form a 32 entry deep hardware queue for requests from the ROS application. The number of available
entries is maintained by the ROS application. To provide the destination addresses, the ROS
application selects a buffer from the memory pool. A ROBIN configuration parameter assures that the
actual fragment size cannot go beyond the size of the buffer area. This is done by setting an upper limit
to the maximum number of fixed-size memory pages the ROBIN may use for any event fragment. Any
fragment exceeding that size is truncated by the ROBIN during reception from the link.
3.3.3.3 ROS Software
The ATLAS TDAQ online software is made up from a very large number of individual packets, which
control trigger, DAQ, database access, monitoring etc. Alone the dataflow section, which is the main
framework for the ROS and the ROBIN consists of 200 software packages. Among these, 27 deal with
the ROS, 2 contain the application and boot code of the ROBIN and 1 covers the ROBIN FPGA code.
Three of the ROS packages are relevant for the ROBIN and contain device drivers, a library and a
number of applications. These packages are currently maintained by the CERN ROS software team
and the group at RHUL. The Mannheim team has been and will be active in this area as well, however
there is no manpower available at this time.
The main device driver performs the initialisation of the kernel structures related to PCI devices and
makes the resources available to user applications. The driver also accepts interrupts from the ROBIN
and provides the hooks for a corresponding user level interrupt handler. Additionally, it provides some
debugging information which is available through the Linux “/proc” filesystem and which reports for
every ROBIN card the serial number, the version number of FPGA and application code, the real-time
status of the S-Link inputs, occupancies of the request queues and the values of the mailbox
communication registers. Apart from the debugging interface the device driver is very generic and
leaves most of the device specific functionality to a user level library and the associated applications.
This approach makes the device driver less sensitive to modifications of the ROBIN functionality,
which is advantageous as installing a new device driver has to be done by the system administrators
while the applications can be updated by regular users.
A second device driver is available which provides a standard serial port to the host. This serial port is
implemented at the hardware level by the the ROBIN FPGA, which in turn attaches to a serial port of
the ROBIN processor. The purpose of this driver is to gain access to a ROBIN terminal port without
attaching a cable. The latter is not practical, as the ROS-PC has only one internal serial port. Changing
cables or adding an external USB-to-serial expander are not viable options under normal operating
conditions. The serial interface is then used for testing and debugging purposes, for example the test
suite (see below) uses this feature to set configuration values and to retrieve status and debug
messages. There are two reasons to keep this driver separate from the main driver. Firstly, the
functionalities are completely different and the serial port is only used for debugging and maintenance.
Secondly, the serial driver interferes with the firmware upgrade procedure and must be unloaded
beforehand. To minimise the chance for a system crash due to this interference, the serial driver is by
default not loaded.
In addition to the two ROBIN device drivers there are other drivers used by the ROS software, for
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
35
Chapter 3 - ATLAS DAQ
example to allocate physically contiguous memory for the communication with the ROBIN.
The library used by the drivers and the applications contains functions related to the following areas:
•
Allocation of boards and resources, memory mapping
•
Access to the PLX PCI-bridge device
•
JTAG access to FPGA and CPLD
•
Exception and interrupt handling
•
Handling of FPGA bitstreams, incl. FPGA configuration
•
Message passing
The application “robinconfigure” is responsible to create the link between the ROBIN cards and the
memory pool used for the response messages. This application is normally called by the device driver
during initialisation, but can be used later on to modify the memory settings. Every ROBIN consumes
approximately 80kB per channel from a pool of 1MB of physically contiguous memory. This large
contiguous memory is obtained from the Linux operating system at boot time via the “bigPhysArea”
patch.
The utility “robinscope” is the main test and debug tool for the regular maintenance and makes the full
functionality of the PCI message passing interface available to an expert user. The configuration
parameters and monitoring values can be retrieved, the configuration can be modified, fragments can
be uploaded to the ROBIN in different emulation modes and subsequently requested. This includes the
generation of incorrectly formatted fragments and the check for proper error handling plus a simple
performance tests using the internal fragment generator.
Further test utilities are “robinTestSuite” and “robinstress”. “RobinTestSuite” is basically a tool for
testing at the factory level.
Figure 22: ROBIN TestSuite
36
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
It steers and monitors the ROBIN BIST procedure via the serial interface (cable or driver) and tests the
network interface with the help of external network requester program. It also interfaces to the lowlevel configuration tool to enable the factory programming. Multiple ROBINs can be processed in
parallel while the results are displayed on a simple GUI (Figure 22). The purpose of “robinstress” is to
request fragments over PCI at the maximum rate while checking the fragment CRC code in order to
verify data integrity. Normally, there shouldn't be a problem with this on PCI but there were a number
of incidents as described in section 6.3 .
The resident firmware of the ROBIN has different sections, which are all contained in a single FLASH
memory device. The tool “robin_firmware_update” allows to access the FLASH. The FPGA firmware
and the application code are updated most frequently. As the versions of the two must match, the tool
updates both of them in simultaneously. The boot code and the low-level environment settings are not
very likely to change and normally need to be written only once after the production. Product data –
serial number, hardware version, production date and site – are written to a special region in the
FLASH device which is one-time-programmable and must be initialised after the factory test.
The main ROS application “ROSApplication” uses the library to set the configuration parameters
according to the values in the global configuration database, but limited to non-expert parameters. The
channels of the ROBIN are enabled or disabled according to the global run-control state of the TDAQ
system. The online monitoring systems requests the operational monitoring data on regular intervals.
Fragments are requested by a multi-threaded request handler, which queues a number of requests per
Figure 23: ROS-ROBIN interaction
channel to the ROBIN. Delete requests can be interleaved into the queue. The interaction between the
ROS application and the ROBIN is depicted in Figure 23.
3.3.4
Switch-based ROS
As explained in chapter 3.3.3 the performance of a standard 12-channel ROS-PC is just above the
standard requirements. If the search for new physics requires to run the TDAQ system at higher rates
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
37
Chapter 3 - ATLAS DAQ
more ROS-PC with fewer ROBINs are required, which consumes eventually much more rack space.
An alternative approach to achieve higher performance with almost the same system density is the
switch-based ROS. Here, the ROBINs are connected to the dataflow network via their private GE
interfaces. Due to the reduced load on the ROS-PC a fifth ROBIN can be installed. The 5 additional
network ports per ROS-PC can be connected to an additional switch per rack, which takes the place of
one of the ROS-PCs. As the total number of channels per rack does not necessarily change, this
implementation has relatively little impact on the overall installation.
A more extreme variant of the switch-based ROS has been investigated based upon the idea of a ROBon-ROD: the ROB replaces the S-Link source card on every ROD and connects directly to the
dataflow network [ROBROD]. This implementation has a very high flexibility and performance but
requires a large number of network ports and significantly complicates commissioning. Although a
procedure to solve the commissioning issues was proposed [ROBRODC] this solution was mainly
dropped due the problems expected in that area, due to the ROBs belonging physically to the ROD
system but logically to the TDAQ system. Nevertheless, the requirement remained on the ROBIN to
be able to prototype such an architecture.
The current view of the switch-based ROS is just a variation of the standard bus-based architecture
and is applied only to sub-detectors which require performance not achievable otherwise. The most
likely scenario is that the ROS-PC will remain responsible for configuration, monitoring and
distribution of delete messages to the ROBINs. Possibly, it will also collect fragments from the
ROBINs for the EB requests. The ROBINs will individually respond to L2 requests via their network
interfaces in a way, that every ROBIN effectively implements a ROS subsystem containing 3
channels. With 5 ROBINs per ROS-PC the available network bandwidth on the L2 network is about 5
times higher than of a standard ROS-PC, which roughly matches the rate-performance ratio of a
ROBIN and a standard ROS-PC.
3.4 Summary
The main parameters of the ATLAS detector are very similar to CMS. A custom L1 stage reduces the
initial event rate (GHz range) to 100kHz. The data corresponding to a full event are generated by 7
sub-detectors, distributed over 1600 sources and pushed into the DAQ/HLT subsystem via optical SLinks. The nominal output to mass storage operates at 100Hz with an event size in the order of 1MB.
The entire ATLAS DAQ/HLT system is build from commodity technology – standard PCs and GE
networks – with the exception of the ROBIN components which establish the interface to the detectors
plus a unit which controls the event assignment – the region-of-interest-builder (RoIB).
The mechanism of the dataflow internal to DAQ/HLT is very different from the traditional “PUSH”
model. Initially, all event fragments are received by the read-out-system (ROS) which provides
buffering for a few hundred ms. The ROS is composed of standard PCs each housing typically 4
ROBINs. About 2% of the stored fragments are pulled from the ROBINs by the L2 trigger subsystem,
which performs a quick analysis based upon regions-of-interest and sequential selection. For events
passing L2 all fragments are pulled by the event builder (EB) subsystem, for full online analysis. Due
to the “PULL” model, all events have to be explicitly deleted from the ROBINs, which is done via
38
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 3 - ATLAS DAQ
broadcasts generated by a central dataflow-manager (DFM).
For the ROS two architecture variants are considered. In the baseline bus-based ROS only the PC
interacts with the rest of the system, forwards the requests to the ROBINs and collects the returning
data. The performance of the PC's processor and memory limits the maximum request rates in this case
to around 10kHz. The enhanced switch-based ROS allows the ROBINs to interact directly to the rest
of the system via its private GE port. In this case, the ROBIN becomes the performance limiting
component, at a request rate around 20kHz.
The size of the system with respect to network connectivity is in the order of 300 (baseline) to 1.000
(enhanced) ports for the ROS, 700 ports for L2 and 300 ports for EB. The final implementation uses
two large central switches plus a number of concentrator switches which group some ROSes and L2
processors respectively. The network components are arranged such that the load under nominal
conditions stays below 60% capacity of the switches, in order to avoid message loss. Simulations and
measurements on large-scale test setups have shown that the required performance can be obtained
with GE.
The ROBIN component is exposed to the high input rate of 100kHz on each of the 3 input channels
and has to perform bookkeeping of up to 64k stored events per channel. On the output interface, it has
to deal with requests from PCI and GE, which can be active concurrently with a combined nominal
rate of 6kHz and a maximum rate of 21kHz per channel. In addition, it has to run complex operational
monitoring tasks. The final ROBIN implementation uses the combination of microprocessor and
reconfigurable logic (FPGA) technologies, providing a cost efficient design tailored to the specific
requirements of ATLAS.
The ATLAS dataflow system does not provide scalability via an inherent granularity as CMS does.
The entire ROS and the central switches must be present in any case, which is a significant constant
offset in terms of resources. On the other hand, HLT performance can be increased virtually in terms
of single machines.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
39
Chapter 3 - ATLAS DAQ
40
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
4
FPGA Technology
FPGA technology addresses application areas which require more performance than software solutions
can provide on one hand but which cannot use custom hardware design with digital logic components
and ASICs on the other hand due to flexibility and cost issues. In principle, an FPGA is a silicon chip
with a large number of simple, uniform logic elements (LE) and a large number of I/O pins. These
elements are used to hard-wire the required functionality, while the re-programmability of the device
allows to update or entirely change the behaviour of the circuitry under user control.
FPGAs are ideally suited to implement any simple logic functions, dataflow and memory controllers
and certain processing algorithms, for example which use primarily simple parallel processing. These
features match very well the requirements of this project. The following paragraphs introduce the basic
elements of FPGA technology, the development tools and some prominent examples of library
elements used in or at least considered for the project, with a focus on the XILINX Virtex-2 device
family used on the ROBIN.
4.1 Device types
FPGA technology was introduced in 1984 by XILINX. Since then, a number of different FPGA types
have been produced by XILINX and other vendors like ALTERA, ATMEL and LATTICE. To date,
there are two main branches in FPGA technology: one-time-programmable (OTP) and volatile. The
OTP branch directly addresses high-volume applications, where true ASICs cannot be used for
whatever reason. Also, radiation hardness is quite good with OTP technology, which makes it the first
choice for space-bound and similar application areas. The volatile branch uses on-chip static RAM to
store configuration data. This requires chip initialisation after every power-on transition but provides
an infinite number of reconfigurations. Due to the flexibility required for the ATLAS ROBIN, only the
volatile technology is viable and OTP technology has never been considered.
4.2 Basic elements
The LEs that build the dominant portion of an FPGA are based upon a small memory and a storage
cell. Typically, the memory has 4 address inputs and a single data bit with separate input and output.
Each of the addresses is an input pin to the LE. The data output can be routed directly or via a storage
cell to the output of the LE. In the VIRTEX-2 [XLNXDS31] FPGA family two LEs are grouped
together to a slice. Additional functionality per slice enables to configure polarities, clock edges, carrychains, dual-port memory usage etc. The arrangement of a single LE is shown in Figure 24.
The logical function of the LE is implemented by an appropriate initialisation of the memory, which is
used as a look-up-table (LUT). Due to the low number of inputs, complex logic requires the cascading
of LEs in order to generate functions depending of many inputs. Alternative to logic functionality the
LE memory can be used as dedicated memory, for which depth (address lines) and width (data bits)
expansion is possible as well. FPGAs can also provide a large number of I/O pins, called IOBs. To
accommodate the use in many different environments I/O voltages and I/O standards are configurable
with voltages in the range from 1V up to 3.3V.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
41
Chapter 4 - FPGA Technology
The Virtex-2 family for example supports 25 different I/O standards including differential ones, plus
programmable drive strength and on-chip termination. Standard I/O cells are capable to run doubledata-rate (DDR) up to 1.25Gbit/s in Virtex-5, and specialized I/O cells up to 6.5Gbit/s. A Virtex-2 or
Virtex-5 chip is divided into several I/O banks. All IOBs in a bank share the same I/O voltage and
access to global resources like clocks and reset signals. I/O standards within a bank can be different
however, as long as the voltage requirements are compatible. For example, LVTTL, LVCMOS, LVDS
Figure 24: Virtex-2 logic element
and SSTL can be used in a single bank running from 2.5V. The package layout and I/O banking
scheme of the FPGA on the ROBIN is shown in Figure 25.To connect the LEs and the IOBs a flexible
routing fabric is required, which is realised via traces implemented in several layers42 of metallisation
and programmable transistors for the connections. All resources – slices, IOBs and special functions –
are arranged in a matrix with row and column organisation. Every row and column has access to
routing resources of different layers:
42 Virtex-2 has 10 metal layers, Virtex-5 has 12 layers.
42
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
•
Long lines
•
Hex lines
•
Double lines
•
Direct connect lines
Figure 25: Virtex-2 FF896 package
For example, direct connect lines connect adjacent LEs and IOBs only while long lines provide
connectivity across the entire device.
The evolution of FPGA technology over the last years is shown in Table 3. In the beginning, FPGAs
were mainly used to implement simple logic functions (glue-logic). With increasing size full
applications – for example image processing algorithms – could be implemented. The addition of
special functions enabled to build complete systems on a chip (SoC).
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
43
Chapter 4 - FPGA Technology
Year
Family
Structure
LEs
System speed
Pins
Special functions
1995
XC4000
250nm
5000
80MHz
500
No
2003
Virtex-2
150nm
20000
150MHz
1500
Some (internal)
2008
Virtex-5
65nm
100000
300MHz
1700 Many (internal, I/O)
Table 3: XILINX FPGA families
For the Virtex-2 family, extra functional blocks were added to the silicon for memory, clock
management and math functions. Moreover in Virtex-5, there are embedded processors, triple-speed
Ethernet controllers and high-speed serial I/O blocks.
4.3 Tools
There are four categories of tools related to FPGA designs which cover the following areas:
•
Specification of the functionality
•
Simulation and test
•
Synthesis
•
Vendor specific tools
and which are all together used in a typical design process.
4.3.1
Specification
There are different methods to specify the functionality of an FPGA. In the early days and with simple
functionality the logic functions of the LE were edited and the connections created manually or via
scripts. Schematic editing tools were used frequently later, which enable to use libraries of standard
TTL functions like multiplexers, counters, flip-flops et cetera to create a design description for an
FPGA just like for a electronic board. However, this method also requires a thorough hardware
expertise from the designer and for large devices the schematic drawings become unmanageably large.
A further step was the application of a standard hardware description language (HDL) like VHDL or
Verilog to FPGA design. The HDL approach enables to create a text based structural description using
the same elements as used for the schematic description. But more important it allows to define
functionality in a procedural way, which resembles the standard software development process, at least
to a certain extent. In VHDL for example, the designer can define data types, variables, constants,
functions and procedures and make use of control structures like if-else, case and while statements. A
typical VHDL code snipped is shown below:
44
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
--------------------------------------------------------- Multiplexer for transmitting data to TLK
-------------------------------------------------------tlk_tx_multiplexer: process (tlk_tx_clk, bypass_cntl_data(0))
begin
If bypass_cntl_data(0) /= '1' Then
tlk_txd
<= hola_txd;
tlk_tx_en
<= hola_tx_en;
tlk_tx_er
<= hola_tx_er;
Elsif Rising_Edge(tlk_tx_clk) Then
If txFifo_data_at_output = '1' Then
tlk_txd
<= txFifo_dout(15 downto 0) after 1 ns;
tlk_tx_en <= txFifo_dout(16) after 1 ns;
tlk_tx_er <= txFifo_dout(17) after 1 ns;
Else
tlk_tx_en <= '0' after 1 ns;
tlk_tx_er <= '0' after 1 ns;
End If;
End If;
end process tlk_tx_multiplexer;
The code starts with comments lines, then a control structure with the name tlk_tx_multiplexer of type
process is defined. The parameter list to the process defines input signals to which the process is
sensitive, here tx_tlk_clk and bypass_cntl_data(0). There are other input signals as well, which are not
on the sensitivity list. The difference is, that the transition of any output of the process occurs only
when a signal from the sensitivity list changes. A transition of a normal input has no immediate effect
on the outputs. The variables used in this code snipped are signals which can normally have one of two
values – '0' or '1' – if seen as electrical signals. However in a typical VHDL description the signals are
logical signals which can have one of multiple values: apart from '0' and '1' there are 'Z' to indicate a
high-impedance state, 'X' to indicate unassigned, 'U' to indicate undefined and 'H' and 'L' to indicate
weak pull-up and pull-down states respectively. The main purpose of this multi-value-logic (MVL) is
to improve simulation of the circuitry. Signals can be grouped to vectors and vectors can be references
to as a whole or by single or multiple elements.
Nevertheless, creating a design specification with VHDL still requires the skills of hardware expert, as
there are many constructs which are not common to software programmers. Other approaches have
been developed to generate a hardware description from “C”-style code 43, enhanced with some
features to specify parallelism, or to integrate FPGA libraries into graphical programming
frameworks44. Such tools make it easier for software experts to create FPGA design specifications but
still lack flexibility and performance compared to the HDL methodology.
4.3.2
Simulation and test
Once a designs specification has been created, the functionality has to be verified using a functional
simulation. For a VHDL based specification, this is done using a VHDL simulator 45. The simulator
creates an executable representation of the design which enables to probe and stimulate any signal in
the design. Typically, the external signals of the design are stimulated by a series of so-called test43 A tool to create FPGA code from a C-style description is “Impulse-C”: www.impulsec.com
44 A common graphical programming framework is “Matlab” (www.matlab.com) which is specialised for DSP
style FPGA applications.
45 A common VHDL/Verilog simulator is “Modelsim” from Mentor (www.mentor.com).
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
45
Chapter 4 - FPGA Technology
vectors, which are applied one after the other. For every test-vector the executable is run until a steady
state is reached, then the outputs are updated. Simple test-vector sets can be created manually.
Complex test-vector sets, for example the access sequence of a microprocessor over a bus, can be
generated via scripts or via external programs. The results of a simulation run (the application of a set
of test-vectors to the simulation executable) are normally viewed as a waveform, as shown in Figure
26.
Simulation has a number of advantages. First, the simulation can monitor internal signals which are
not visible on the physical boundary of the device. Next, creating the functional simulation executable
can be done much quicker than creating the configuration data-set. Finally, simulation can be done
without having the target hardware available. However, there are disadvantage as well:
•
The precise timing behaviour of the design is not properly considered, which can lead to a
mismatch between the results of the functional simulation and the real design.
•
Simulation is normally much slower than real operation, for example simulating 1ms of real
time can require 1min of simulation time or more. To achieve acceptable simulation times the
source designs have frequently to be modified46 for simulation purposes.
Figure 26: Simulation waveform
To compensate the potential timing mismatch modern VHDL simulators are able to use vendor
supplied simulation models, which provide information of inherent signals delays related to the basic
FPGA elements. Signal delays introduced by signal routing however are basically ignored. A precise
timing model – including all routing delays – of a design can be obtained after the entire design
compilation has been completed. However, the execution time of a simulation using a timing accurate
46 For example, an I/O controller requires stable input signals for initialisation during 1ms of real time. For
simulation purposes this timing requirement could be shortened to 10µs.
46
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
model will normally be higher by a one or two orders of magnitude, which limits the usefulness to
very specific situations.
Testing of the FPGA functionality in real life is on the first view limited to the observation of the I/O
signals via an oscilloscope or a logic analyser. For complex FPGA designs this is totally insufficient.
Tools to access internal signals in a way similar to accessing external signals are available from FPGA
and synthesis tool vendors and build upon the feature of the FPGAs that the internal resources can be
monitored and controlled via a JTAG interface. The typical approach followed by a tool like
ChipScope47 is to use some internal FPGA memory to store samples from a selection of interesting
signals. The insertion of memory, controllers and of additional signal connections is done during the
synthesis step. Trigger conditions can be defined and uploaded via JTAG at run time. The results are
read via JTAG and displayed in a waveform view (see Figure 27), similar to the simulation view.
Figure 27: ChipScope waveform view
This test mechanism enables a deep insight into the internal functionality during run-time. Compared
to the simulation the number of samples is very limited (a few thousands) and every modification of
the signal set for sampling or triggering requires to re-run the time-consuming physical compilation.
4.3.3
Synthesis
The synthesis step starts at the source of the specification, just like the simulation. For a structural
description based upon libraries made from basic elements this is a straightforward process and done
simply by expansion of the libraries to the network of basic elements. However an abstract HDL
description requires a complex tool to analyse the description and to create the proper low-level
description suitable for the FPGA. The statements of the HDL source must be translated into a
network of basic FPGA elements, for example into flip-flop registers with asynchronous reset like in
the code snippet above, or into embedded memory structures. This process is very specific for
47 ChipScope is the test tool from XILINX.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
47
Chapter 4 - FPGA Technology
different FPGA families and vendors. Advanced synthesis tools48 utilise timing information already at
this step and optimise the generated logic according to various user constraints, e.g. they minimise
delay or minimise space. The result is then a fully mapped structural description, where the basic
elements are already grouped together according to the physical resource constraints49 of the target
FPGA. The performance of state-machines, arithmetic statements, counters and complex control
structures depends strongly on the quality of this translation and sometimes a performance goal can
only be met when the design is synthesised with a special tool.
4.3.4
Vendor specific tools
The network of basic elements created by the synthesis must be distributed across the resources
available on the chip and the proper connectivity must be established. This process is called place-androute (P&R). If the mapping has not already been done the synthesis tool, the logic functions and
registers must be grouped to match the multiplicity of the physical FPGA structure. Once the P&R is
complete, a timing analysis has to be performed to verify that the timing constraints are met. Timing
constraints are defined for example by specifying the clock period for every clock network in the
design and by “input setup to clock” and “clock to output” parameters. If the timing goals are met, the
implementation behaviour should be consistent with the simulated behaviour. If the goals are not met,
different settings of the P&R tool may improve the results, or the source design has to be modified. In
many cases, the timing analysis provides a good indication which signal paths have to be improved. A
typical modification is to add one or more register stages (pipeline registers) to a construct which
requires many logic levels50. The final step after P&R is the generation of the configuration data-set –
the bit-stream – which has to be downloaded into the device. During the bit-stream generation also a
final design-rule-check is performed. A major task of the bit-stream generation program is to perform
the mapping of configurable elements to bits in a way that prevents reverse engineering attempts. In
newer devices there is also the option to encrypt the bit-stream. The decryption is done on-chip with a
key stored in a battery-buffered memory.
4.4 IP-Cores
All FPGA vendors provide libraries of special functions which help implementing complex designs.
Some of these library elements are available as HDL source code, but most of them are available in the
form of encrypted netlists, which can be added via a wrapper-construct to the HDL or schematic
specification. The netlist is then included by the vendor tool-chain after the synthesis step. There are
also third-party companies which develop and sell such modules as intellectual property (IP) cores.
Examples for such IP cores are processors, communication controllers like PCI or Ethernet,
encryption/decryption modules, etc. Recent FPGAs also provide a number of embedded51 cores like
block-memories, multipliers, processors, Ethernet MACs and serial Gigabit transceivers. The
48 Advanced synthesis tools are for example “Synplify Pro” from Synplicity/Cadence (www.synplicity.com) or
“Precision” from Mentor (www.mentor.com).
49 This grouping takes into account how functions and registers can be allocated to a slice of LEs
50 A logic level corresponds to a single LE. Multiple logic levels require cascading of LEs, which increases the
delay.
51 Embedded cores are hardware modules on-chip, which do not consume FPGA resources.
48
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
synthesisable cores are commonly called soft-IP, the embedded cores hard-IP. The ROBIN makes use
of both soft and hard IP. Hard IP is used in the form of memories blocks which implement dual-port
memory and FIFOs. The Gigabit-Ethernet MAC is an example for a soft-IP core. The external
processor of the ROBIN was selected to be of the same kind as the embedded processor of more recent
XILINX FPGAs, in order to facilitate a potential migration to a denser implementation.
4.4.1
Embedded processors
Common embedded processors (hard-IP) are based on either ARM (ALTERA) or PowerPC (XILINX)
technology. XILINX Virtex-2Pro and Virtex-4FX FPGAs include one or two PowerPC-405 cores
which can operate at up to 400 or 450MHz respectively. XILINX Virtex-5FX FPGAs include one or
two PowerPC-440 cores which can operate at up to 550MHz. Using the embedded processors one can
implement a complete system-on-chip (SoC), equivalent to a micro-controller with customised
peripherals. The interface between PowerPC core and FPGA fabric runs via the standard PowerPC
processor local bus (PLB), which translates between the PowerPC core frequency and the clock
domain of the FPGA logic. The embedded processor cores can be interfaced to internal blockmemories quite easily but adding external memory consumes a significant amount of FPGA resources,
at least for the Virtex-2Pro and Virtex-4FX FPGAs. Communication between the processor core and
FPGA logic can be done via different interfaces, of particular interest are the “fast simplex link” (FSL)
ports and the “auxiliary processor unit” (APU) interface, which are both supported by special
processor commands. Both interfaces should provide very efficient communication mechanisms.
The XILINX tool suite “embedded development kit” (EDK) allows to develop processor based FPGA
designs, where the processor(s) and a number of IP-cores can be combined to form an embedded
system. Custom user IP-cores can be developed following an IP-reference design and added to the
system. Programming is done using the open-source GNU52 tool chain. The same tool-suite can be
used for systems based on the XILINX soft-IP processor “MicroBlaze”, which has a 32bit RISC
architecture similar to the PowerPC operating at 60 to 125 MHz, depending on the FPGA platform.
4.4.2
Ethernet MAC
Ethernet is a very common communication standard and is also widely used in embedded systems.
Typically, an Ethernet interface is built from two devices, a physical interface adapter (PHY) and a
media access controller (MAC). The PHY implements layer 1 of the OSI layer model [OSI] and
attaches to the physical media, which can be optical or electrical. Common speed-grades are 10, 100
and 1000Mbit/s, 10Gbit/s is about to move into the commodity market, higher speeds are under
development. The electrical media normally uses pulse-transformers to provide electrical isolation.
The electrical media uses 1, 2 or 4 signal pairs in an RJ45 connector53. The interface between PHY and
MAC follows one of the “media independent interface” standards – MII, RMII, GMII, RGMII or
SGMII – depending on the speed and the device type 54. There is an additional control interface to
52 GNU software is available from www.gnu.org.
53 The RJ45 connector allows full-duplex operation, if the link partners are peer-to-peer or connected via
switches. There are other (older) technologies as well, which are not relevant for this work.
54 MII is for 10 and 100Mbit/s, the “G” types are for 1Gbit/s, “R” indicates dual-edge signalling, the “S”
indicates serial interface.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
49
Chapter 4 - FPGA Technology
access the registers in the PHY, implemented as a bi-directional serial bus.
The purpose of the PHY device is to translate between media signal levels and standard logic levels
and to perform the negotiation of the capabilities55 between the link partners. The MAC device
implements layer 2 of the OSI model and runs the Ethernet protocol [ETH802.3]. Ethernet uses
packets with a fixed structure:
•
A 22 byte header56
•
Payload, maximum57 1492 bytes
•
Padding up to the minimum packet size of 64 bytes, if required
•
A 4 byte frame-check-sequence (FCS)
On the ROBIN a 1Gbit/s MAC soft-IP (GMAC) is used to implement the MAC functionality and
connects to the external PHY via the GMII interface, which has 8 data bits plus 2 control bits per
direction, running at 125MHz. On the user side the GMAC implements two independent FIFObuffered ports, one for transmission and one for reception. The ports are 32 bit wide, plus control bits
to indicate start and end of packet. In addition the GMAC collects the standard network monitoring
values like number of transmitted/received packets and TX/RX errors, which can be retrieved from a
set of registers to facilitate debugging in case of communication errors. A drawback of this soft-IP is
that is does not support multiple speeds, hence the ROBIN cannot communicate with link partners of
lower speeds. A possible work-around would be to additionally implement a MAC which supports
10/100Mbit/s communication and to switch between the MAC cores according to the negotiation
result of the PHY. However, this would require an prohibitively large amount of FPGA resources and
is not necessary in the normal working environment of the ROBIN.
New FPGAs like Virtex-4 and Virtex-5 include multiple triple-speed (10/100/1000Mbit/s) hard-IP
Ethernet MACs, which consume very little additional FPGA resources.
4.5 Summary
XILINX Virtex-2 FPGA technology as used on the ROBIN allows to run at system speeds of up to
150MHz, several hundred pins are available for user I/O plus several complex functional block like
configurable memory arrays and DSP blocks. Functionality is realised via a bit pattern (so-called
bitstream) which controls the operation of a large number of simple processing elements and I/O cells.
The bitstream is stored in volatile on-chip memory, which allows an infinite number of configuration
cycles. The configuration itself is created by a set of tools in a multi-step process which starts with the
synthesis of a higher-level description – frequently called a design – into logic expressions – this is the
RTL description. Pre-compiled library elements (so called IP-cores) can be used at this stage as well,
which implement complex functions, for example a GE-MAC, a processor or a filter operator for
image processing. Next, the equations have to mapped onto the basic elements of the device. Then the
basics elements must be placed on the chip and the interconnects must be established. Finally, the
55 Link partners with different speed negotiate for the fastest common speed.
56 According to IEEE 802.3
57 The standard allows to go beyond the maximum packets size using “jumbo” frames of up to 16kB.
50
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 4 - FPGA Technology
bitstream is generated which contains the configuration value for every single programmable element
of the FPGA. The bitstream must be loaded into the device after power-up, typically from an external
non-volatile memory, via a JTAG interface or from a microprocessor. A typical functional module 58 of
an FPGA designs consists of many LE and requires to run at a specific operation frequency. The most
critical task of the tools is to arrange the LEs on the chip in a way, that all interconnects are fully
routable and the signal delays introduced by the routing are within the timing specification. This
requirement is called timing-closure.
58 For example a counter or a basic computation of an image processing task.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
51
Chapter 4 - FPGA Technology
52
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
5
ROBIN
This chapter presents the implementation of the ROBIN. It starts with a summary of the requirements,
both in terms of performance and functionality, followed by a description of how the ROBIN project
was managed from the early development stage through prototyping and volume production up to full
operation. The basic hardware, the FPGA code and the processor application code are presented in
detail. The chapter concludes with a description of the installation and commissioning procedures.
5.1 Requirements
The functionality of the ROBIN comprises different task areas. The most important group covers the
very basic aspects of data handling – input and output – and the related bookkeeping. The key
parameters (see section 3.3 ) are 100kHz event rate, 160MB/s input bandwidth, request rate of 21kHz,
buffer latency in the order of 100ms and a multiplicity of 3 channels. The total nominal output
bandwidth is relatively low, in the order of 100MB/s. To support both bus-based and switch-based
ROS architectures a PCI interface and a private GE interface are required. The bookkeeping
mechanism keeps track of the position of individual event fragments in the buffer memory. This
information is needed to retrieve the data and to de-allocate the memory space once the event is
processed.
Next, message handling functionality of various complexity is required in order to make use of the
data handling. For the bus-based ROS, sending requests and receiving responses is straightforward, as
the ROBIN is a standard PCI peripheral to the ROS-PC, which means peer-to-peer communication
over a reliable media. In switch-based mode the ROBIN has to maintain network connections to
several hundreds of nodes in parallel, running an unreliable protocol (UDP) over GE. Initially, the
interfaces PCI and GE were meant to be operational mutually exclusive. However the planned
approach to maximise the request rates is to use both concurrently.
Furthermore, ATLAS requires operational monitoring capabilities, like statistics of fragments,
messages, buffer occupancies and errors. Initially, the main issues here were covered by simple
counting of the associated occurrence (incoming fragments, for example). During the installation and
commissioning phase of the ATLAS DAQ system however the monitoring was expanded significantly,
for example by classification of transmission errors, a recording facility for corrupted fragments which
do not follow the standard event format and an interrupt capability towards the host PC.
Finally, a set of functions must be available related to setup and configuration. The run-time behaviour
is controlled via a number of configuration parameters. A small set of parameters controls the basic
operation by setting the input mode and providing a tag which globally identifies every channel. Some
parameters are related to the buffer management, like size and number of memory pages and input
truncation threshold. A couple of parameters influence very particular behaviour and are used only
during debugging. The static setup – FPGA bitstream, boot loader, application binary and default
configuration parameters – is contained in a FLASH memory. The FLASH memory can be modified
from the host PC via the PCI bus and by the ROBIN itself.
Besides the functionality listed a above, an important requirement on the ROBIN is flexibility. From a
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
53
Chapter 5 - ROBIN
system perspective, the majority of the tasks are shared between the ROS-PC and the ROBIN in the
bus-based architecture. Even for the basic dataflow the PC participates, by interfacing to the network
and by combining the fragments from the individual channels into larger packets. In the switch-based
architecture the ROBIN takes over the entire dataflow and even assembles the fragments from its
channels prior to sending them off to the network, but configuration and monitoring are still handled in
cooperation with the ROS-PC. In the most extreme scenario however – and the ROBIN was designed
to prototype even that – the ROBIN operates fully in stand-alone59 mode and still needs to provide the
complete functionality. All configuration and setup – including firmware upgrade – has then to be
handled by the ROBIN itself, controlled by a remote note over the GE interface.
The following sections describe the development and production process as well as the
implementation of hardware and software of the ROBIN.
5.2 Project Management
As mentioned earlier there were already several prototype implementations existing prior to the
current ROBIN design and the prototype results provided guidelines [ROBSUM] for the final
development. These early prototypes were built by different groups60 of the ATLAS TDAQ
collaboration which used different form factors like PCI or PMC, different FPGAs, local processors –
if any – and different buffer memory technologies like SRAM, ZBT and DRAM. The different
approaches are explained in more detail in section 5.3 . After an initial analysis phase a small design
team was formed, consisting of members from three institutes61, with the mandate to propose a multichannel ROBIN solution. The initial requirements were updated with respect to performance,
functionality, space and cost. Concerning the hardware development it was decided to first build a new
prototype62 supporting two input channels and able to demonstrate all possible options with respect to
the system architecture, even two different GE media interfaces – an optical and an electrical one. The
final ROBIN should then be derived from the prototype-ROBIN, by leaving out unused functionality
and eventually adjusting channel multiplicity and mechanical format. The design team – led by the
author of this thesis – produced the prototype design proposal in the form of three design documents
[HLDDPROT][DLDDPROT][SWIDPROT] which were reviewed and approved by a CERN expert
group [FDRPROT]. Subsequently the prototype ROBIN schematics were created at Mannheim,
followed by the specification of design rules for the PCB layout, which was done by an external
company. Production and assembly of the initial cards was organised by the Mannheim group again,
meanwhile the initial FPGA firmware and test software was produced at the three institutes and
integrated afterwards. The location of the main developers at different sites and even different
countries was a serious complication during debugging phases, in particular after the production of
new initial cards. Video-conferencing tools were not satisfactory for detailed technical discussions,
59 A potential implementation would be to install a (eventually larger) number of ROBINs in a housing with a
passive PCI backplane, which just provides the power to the boards. No host CPU would be required.
Alternatively, one of the ROBINs could be configured as a PCI master and act as the host for configuration
and monitoring.
60 Prototypes were build by CERN, Saclay/F, NIKHEF/NL, RHUL/UK and Mannheim/D.
61 Apart from the author there was one person each from NIKHEF/NL and RHUL/UK.
62 This was the so called Prototype-ROBIN.
54
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
hence personal meetings approximately every 6 weeks were held during hardware development phases
in addition to the regular weekly telephone meetings.
In total63, there were up to three persons working on the hardware design, five on the FPGA code and
five on software in the different areas – boot and application code on the ROBIN, driver, library and
test software on the ROS-PC, not counting the application software on the ROS-PC which was
developed by the CERN ROS team. The prototype ROBIN modules were successfully tested and
performance measurements provided good results. As a consequence, the design of the final ROBIN
was – as expected – based on the prototype ROBIN and presented to the ROS community in another
design document [ROBRDD], followed by another review at CERN [ROBFDR]. The preparation of
schematics, layout and initial cards was executed in the same way – which basically means by the
author – as for the prototype ROBIN. After the successful testing of the initial cards a mini-series of
10 cards was produced by the German production company who had done the initial cards as well.
This series was tested not only for functionality and performance but also for reliability, using a
climate chamber at NIKHEF and a programmable power-supply at Mannheim. After these tests the
final stage of the CERN approval mechanism – the production readiness review [ROBPRR] – was
prepared with documents on the final design [PRRDD] and performance [PRRPM], the test strategy
[PRRTP] during production and for the regular self-test and a schedule for the volume production
[PRRPS]. The volume production of 650 ROBIN cards, including some spare cards as specified in the
policy for spares [ROBSPARE], was shared between two production sites64: the German company and
a company in the UK, who used the German production documents to manufacture the printed circuit
boards, to order the components and to assemble the cards. The supervision of the companies was
done locally, in Germany by the author and in the UK by members of the UK group. All cards from
the volume production were tested at the production sites by the local groups using a customized testsuite on top of the built-in self-test functions, which included the external interfaces PCI, GE and
optical input, the latter via a loopback fibre. Sample cards were exposed to the same procedure for an
extended period (typically over night) at the supervising institutes. Installation and commissioning of
the ROBINs was performed primarily by the CERN team, with help from members of the design team.
Finally, maintenance is organised in a way that non-experts at CERN are able to remove defective or
suspective ROBINs from the system, after collecting as much information as possible using an
analysis tool running on the ROS-PC. Those cards are then subsequently checked and repaired (if
possible) by the hardware designers.
5.3 Implementation
The ATLAS ROS system (chapter 3.3 ) allows for different implementation variants of its elements
which today are the ROC-PC and the ROBIN. The two existing architectures - bus-based ROS, which
is the base line and comes historically first, and switch based ROS – differ in the way to concentrate
the data. In the bus-based ROS this is done by a host computer, in the switch-based ROS by a network
switch. Concerning the I/O sub-unit several options have been considered. For a bus-based
63 Most of the developers had particular skills – hardware, FPGA or software – bus a few worked in both or all
areas.
64 Apart from the benefit of having a second source for the cards, the second reason for the sharing was related
to financial arguments.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
55
Chapter 5 - ROBIN
architecture the simplest solution would be to push all incoming data into the main memory of the host
PC (the ALICE way, section 2.1.2 ) and to leave it to the main CPU to handle the events. The CERN
prototype devices SSPCI65 and S32P6466 were used in such a setup as so-called software-ROBINs
[SWROB]. The disadvantage of this approach is the very high load on the memory bus of the host PC
and the high rate of events to be handled by the main processor. As a result, the link density per PC
would be low. Also, this is not an option to implement the switch-based architecture. Solutions to store
the data directly into the host memory while doing the bookkeeping locally on the I/O card were
looked at, but not actually tested, because they solved only the memory bandwidth problem but did not
support the switch-based approach either. In [MMROB] the first multi-channel prototype with local
buffering and bookkeeping was described, using the general purpose PCI-based FPGA co-processor
MPRACE-1.
A step towards the switch-based ROS was made with an S-Link module [GBELSC] using GE as the
output media, which stimulated the investigation of a ROB-on-ROD mezzanine [ROBROD] and the
associated architecture. This approach initially looks simple and cost-effective but comes with serious
complications during commissioning and maintenance. The problem here is that the ROD belongs to
the detector subsystem while the ROB belongs to the DAQ and mixing the two on a single module
makes is very difficult to independently verify the functionality. This is a problem in particular for the
DAQ, as every intervention on a ROB-on-ROD module would require to power off or on the hosting
ROD unit, unless a remote power supply via power over Ethernet (PoE [POE]) is established – which
would in turn require a more complex implementation due to multiple power supplies. Another switchbased option proposed to implement the ROBIN as a mezzanine for a VME [VME] card, with a single
S-Link input and a FE output per mezzanine and approximately 6 mezzanines per VME card. The
VME host would take over the configuration functionality.
An extensive evaluation of all the options led to the implementation of the present ROBIN, which has
a very universal design and basically a superset of the functionality of all investigated choices.
5.3.1
Hardware
As the hardware of the final ROBIN and the prototype-ROBIN are very similar, this section will focus
on the final ROBIN and make references to the prototype-ROBIN where appropriate.
The following basic elements are essential for the hardware implementation of the ROBIN:
•
Multiple ROL/S-Link channels
•
Buffer memory
•
FPGA
•
CPU
•
PCI Interface
•
Network Interface (GE)
65 http://hsi.web.cern.ch/HSI/S-LINK/devices/slink-pci/
66 http://hsi.web.cern.ch/HSI/S-LINK/devices/s32pci64/
56
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
The FPGA is the central unit and all other elements are grouped around it.
Figure 28: ROBIN basic hardware elements
The arrangement in Figure 28 satisfies the conclusions from the guidelines. The CPU is taken off the
main data path, which is handled by the FPGA alone. The buffer management scheme (see 5.3.3.2 )
makes it straightforward to place the management memory at the processor, and in fact to make it part
of the processor’s main memory. The existence of both PCI and GE interfaces was initially required to
investigate the two different ROS architectures – bus-based (see 3.3.3 ) and switch-based (see 3.3.4 ) –
and they were assumed to be mutually exclusive. Later on, the GE interface was considered to be an
additional path, operating concurrently with the PCI interface in certain conditions. The number of
input channels was set to 3 for the final ROBIN and to 2 for the prototype-ROBIN. The MPRACE-1
board (see 8.1.1 ) was used as a template with respect to power-supply67, PCI-interface68 and FPGA
family69. In addition MPRACE-1 was used to rapid-prototype some important functions prior to the
implementation on the ROBIN, by building mezzanine modules for CPU and GE interfaces of both
prototype ROBIN and final ROBIN.
Apart from the obvious requirements in terms of bandwidth and rate for the various interconnects of
the data-path the communication required for the control-paths had to be analysed. There are 2 major
interaction types to be mentioned:
•
Buffer management related messages between FPGA and CPU. An information record of 16
67 MPRACE-1 uses a single 3.3V power source, which is supplied via the PCI connector or optionally via a
ATX-style connector. All local voltages are generated from this main supply, mainly using switching
regulators.
68 A PLX PCI9656 device is used. It bridges a 64bit/66MHz PCI2.2 bus to a 32bit/66MHz local bus interface.
69 XILINX Virtex-2
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
57
Chapter 5 - ROBIN
byte size is generated per page and deposited into a FIFO inside the FPGA, from where it has
to be retrieved by the CPU. The nominal bandwidth required is 1.6 MB/s per channel.
•
TDAQ request related messages. TDAQ requests typically involve multiple messages. First, a
request message is being sent via the TDAQ-interface PCI or GE to the FPGA. Next, that
message has to be retrieved by the CPU from the FPGA. Finally, a response message is sent
from the CPU to the FPGA, which triggers the data transfer over the TDAQ-interface. A PCI
data request requires 20 byte and a PCI delete-request 4 byte per event on average. On GE,
data requests are approximately 4 times larger than on PCI, while the delete-requests have a
similar size. At the nominal 20% request rate the bandwidth is approximately 0.8 MB/s per
channel on PCI or 2 MB/s on GE respectively. The response message sizes – sent from CPU
to FPGA – are in the order of 60 byte for PCI and 160 byte for GE, corresponding to
bandwidths of 1.2 MB/s and 3 MB/s per channel.
Both CPU and PCI-interface provide 32bit/66MHz local buses which are used to connect to the
FPGA. The GE-interface is connected via a standard GMII-interface 70. Therefore, control messages
consume only a small fraction of the maximum bandwidth of the various interfaces involved.
5.3.1.1 FPGA
The selection of the FPGA device family as XILINX Virtex-2 was an initial condition for the ROBIN.
The main remaining parameters to select the particular device are logic resources, memory resources,
clock resources and connectivity. Logic resources are needed to implement control functionality, for
example to handle the HOLA-protocol or to steer data transfers. From the previous prototypes it was
known that the requirements in this area are moderate. Memory resources are used in two different
flavours: as DPR and as FIFO. FIFOs are typically used to transport short messages between the
FPGA and the CPU, like fragment information or DMA descriptors. In case of longer messages, the
FIFO contains message descriptors and the related data reside in a corresponding DPR. This approach
is used for incoming PCI and GE messages.
70 GMII is 8 bit, 125 MHz. In contrast, the protoype-ROBIN had an external MAC connected to the FPGA via
another 32 bit / 66 MHz local bus.
58
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
Memory item
HOLA Core FIFO
S-Link Handler FIFO
Test Input FIFO
LDC Bypass TX
FIFO
LDC Bypass RX
FIFO
Buffer Input FIFO
Buffer Output FIFO
Free Page FIFO
Used Page FIFO
Lbus Header FIFO
Lbus DPR
Lbus Async FIFO
Lbus Request FIFO
MAC Header FIFO
MAC Descriptor
FIFO
MAC Dual Port
Memory
MAC Transmitter
DPR
MAC Receive FIFO
MAC status DPR
PPC Uart
Sum:
BRAMs71
3
3
3
3
Purpose
Buffer and clock decoupling
Flow control buffer
Buffer test event fragments
Buffer TLK transmit data
Size [Bit]
3*512*33
3*256*34
3*512*34
3*2k*18
Buffer TLK receive data
3*2k*18
3
Burst buffer dual-port emulation
Burst buffer for dual-port emulation
Free page list for Buffer Manager
Used page descriptors from Buffer Manager
Header and DMA descriptor for outgoing
message
Buffer request messages from Lbus
Buffer data to Lbus interface
Request descriptor FIFO
Header and DMA descriptor for outgoing
message
Buffer descriptors of received messages
3*512*33
3*1k*32
3* 1K*16
2 * 512*128
512*32
3
6
3
12
1
2k*32
32*32
32*32
512*32
4
0
0
1
512*32
1
Buffer received messages (external memory)
2M*32
0
Decouples clock domains in MAC interface
2*2K*8
2
Decouples clock domains in MAC interface
Ethernet statistic
Buffer data to host PC
15*34
2k*32
2k*8
0
2
1
51
Table 4: ROBIN FPGA memory utilisation
The various memory elements used in the ROBIN are shown in Table 4.
Peripheral components frequently send and/or receive data at their own frequency, which requires
mechanisms to synchronise or decouple various clock domains in the FPGA. The Virtex-2 family
provides DCMs (digitally controlled clock manager) to phase-synchronise internal clock signals with
external clock signals and FIFO-elements with independent clocks to cross clock domains. A total of 8
different clock signals can be used globally72 in the FPGA. The following clock domains are used in
the ROBIN:
•
ROL transmit clock (1): 100 MHz, used to send data to the TLK2501 devices.
•
ROL receive clocks (3): 100 MHz, used to receive data from the TLK2501 devices.
•
CPU clock (1): 66 MHz, used as the common “system” clock, e.g. as the main internal clock
and on the local interconnects FPGA – CPU and FPGA – PCI bridge.
•
Buffer and network transmit clock (1): 125 MHz, used to transmit data to the GE interface and
71 A value of 0 indicates implementation externally or in distributed memory.
72 There are 16 global clock available on chip, but not all of them can be used in all regions concurrently
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
59
Chapter 5 - ROBIN
for all buffer transactions.
•
Network receive clock (1): 125 MHz, used to receive data from the GE interface.
The last global clock resource is used as a global reset signal. The use of DCMs to phase-align the
external buffer clock or CPU clock to the corresponding internal clocks is possible but not actually
required.
As already seen from Figure 28 there are numerous external components connected to the FPGA. On
the ROBIN, all these components use a common electrical standard – 3.3V CMOS – which simplifies
the signal distribution on the FPGA pins. The output signals towards the TLK2501 and the GE PHY
use the XILINX DCI73 feature to provide 50Ohm on-chip serial termination.
Interconnect
CPU
ROL (embedded HOLA S-Link)
Local Bus to PCI bridge (multiplexed A/D)
GE PHY
Buffer Memory (32 bit wide)
MAC External DPR
Test Connector and S-Link Leds
Miscellaneous
DCI control pins
Sum:
Pins per unit
98
44
52
27
57
55
50
12
16
Total number of Pins
98
132
52
27
171
55
50
12
16
613
Table 5: FPGA connectivity
The selected choice74 for the FPGA is a XILINX Virtex-2 XC2V2000-5FF896. With 56 memory
blocks and 624 I/O pins the requirements are met. A large fraction of the available pins are used by the
ROBIN design, the remaining ones being routed to a dedicated test-connector, to facilitate hardware
testing.
5.3.1.2 ROL/S-Link
The S-Link standard primarily specifies a protocol, plus electrical and mechanical properties, but not
the physical transmission layer. In order to implement multiple S-Link channels on a standard PCI
card a particular transmission layer had to be selected for an embedded implementation. Fortunately, a
de-facto standard for the transmission layer – the HOLA75 – was already established when the ROBIN
development started. HOLA is based on optical transmission with bi-directional optical transceivers 76,
at a rate of up to 2.5GBit/s. The SerDes (serialise/de-serialise) function is performed by a TexasInstruments TLK250177 device, with separate 16 bit data paths for sending and receiving. IP-cores for
FPGAs are available from CERN to handle the HOLA protocol for senders (LSC) and receivers
73 DCI means digitally controlled impedance, a features which allows to match the output impedance of
selected IOBs to an external reference resistor, typically 50Ohm.
74 The prototype-ROBIN uses a XC2V1500-4FF896
75 http://hsi.web.cern.ch/HSI/s-link/devices/hola/
76 http://www.finisar.com/download_6ZZvZqFTLF8519P2xNLSpecRevJ.pdf
77 http://focus.ti.com/docs/prod/folders/print/tlk2501.html
60
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
(LDC), which attach to the TLK2501 via the I/O-pins of the FPGA and provide an S-Link protocol
interface to the user application. One LDC-core consumes 1 BRAM and 500 slices in an XILINX
Virtex-2 FPGA. 43 I/O pins are required, including control signals for the TLK2501 and the optical
transceiver. The maximum bandwidth of 160 MB/s as specified by the S-Link standard does not
require the maximum line rate of the optical link, therefore the rate was lowered to 2Gbit/s in order to
improved the robustness. On the user side the nominal clock rate is 40MHz (160 MB/s divided by 4,
for a 32 bit interface) but the HOLA cores can be operated at higher clock speeds, provided the FPGA
timing constraints are met. On the ROBIN, the HOLA user clock is 66MHz, which means that the
maximum user bandwidth is higher than the link bandwidth. Figure 29 shows the connections between
the FPGA and three78 S-Link channels. A 100 MHz clock oscillator provides the common transmit
clock for all channels. On the receive path, every TLK2501 generates its private receive clock,
therefore four clock domains have to be handled inside the FPGA. Control functionality comprises
enable, loop-back and test mode for the TLK2501 plus enable and signal-detect of the optical
transceiver. The serial signals between TLK2501 and SFPs are implemented as 100 Ohm differential
pair transmission lines. The parallel signals between TLK2501 and FPGA are length-matched 50 Ohm
traces. On the receiving side series termination resistors located at the TLK2501 are used. On the
transmitting side the XILINX DCI-feature is used to adjust driver impedance.
Figure 29: ROBIN S-Link interface
5.3.1.3 Buffer memory
The main criteria for the selection of the buffer memory are bandwidth and system latency. With
respect to bandwidth the buffer must sustain 160 MB/s incoming data (the maximum S-Link
bandwidth) plus 20% for outgoing data or roughly 200 MB/s. The worst case is defined by 100%
request ratio, corresponding to 320MB/s. The relevant latency is determined by the processing time of
the L2, which was initially assumed to be in the order of several 10ms (newer estimate is
approximately 100ms, see Figure 18). From these figures the initial buffer parameters are defined: 200
78 The prototype-ROBIN has only 2 S-Link channels.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
61
Chapter 5 - ROBIN
MB/s bandwidth, 1.6 MB size (for 10 ms latency). These values correspond very well to the values
used on the older prototypes and numerous solutions79 are viable. However in a switch-based system
latency can be much larger, due to delays introduced by the network communication or by handling of
eventual message loss. Hence a safety factor in the order of 10 was proposed at least for the prototypeROBIN. A buffer size of 16 MB can sensibly only be implemented with dynamic memory, as the
density of asynchronous SRAM devices is too low and the cost of synchronous SRAM is too high. A
cheap and simple 2 chip80 DRAM running at 125 MHz is used on the final ROBIN 81 to provide 64 MB
of buffer space and 450 MB/s nominal bandwidth.
Figure 30 shows the effective bandwidths for 2 kinds of memory at two different operating
frequencies. The SRAM has a fixed latency per burst of 2 cycles while the dynamic memory has a
frequency dependent latency (due to the fixed timing parameters) and a penalty introduced by the
periodic refresh. It is obvious that bandwidth drops rapidly when the burst-length is shorter than 64
words, which is equivalent to 256 bytes on the ROBIN. Due to the format of data coming from the
ATLAS detectors there is no problem to keep burst-sizes sufficiently high, as the nominal conditions
require fragments of 1.6kB for maximum bandwidth.
Figure 30: Burst-size dependant memory bandwidth
79 Virtually all available memory devices could be used to implement such a system, from simple asynchronous
SRAM to high density DRAM
80 Devices are MICRON MT48LC16M16A2TG, 256MBit, 16M*16
81 The buffer memory on the prototype-ROBIN runs at 100 MHz only, due to the slower speed-grade of the
FPGA.
62
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
The 80MHz used by some older prototypes are not sufficient to sustain concurrent input and output at
nominal speed of the input link, which sums up to 320MB/s. In reality, the instantaneous bandwidth
requirements can be higher, as the fragments are transferred at up to 200MB/s into the memory while a
read access can take place at up to 264MB/s. Burst sizes of 64 words or larger can be easily handled
by the FPGA, hence the 125MHz provide sufficient sustained bandwidth even for 100% readout at full
input rate. The advantage in density of the dynamic memory more than compensates the higher
bandwidth of static memory. All DRAM chips of the three channels are driven by a common 125MHz
clock source, which is also used to drive the transmit section of the GE interface, hence they share the
same clock domain in the FPGA. Per channel 57 I/O pins are required on the FPGA.
5.3.1.4 CPU
From the early ROB prototyping studies the minimum performance of the ROBIN processor for the
buffer management of a single ROL, which is the major task for the processor, was estimated 82 to be
around 60MIPS. A brief survey of micro-controllers has identified two appropriate families, IBM
PowerPC4xx83 and Intel XScale. Apart from a bus-interface to attach to the FPGA and a separate
memory controller no peripherals are essentially required with the CPU core. Both IBM and Intel
provide such simple devices84, in addition to the fully-featured members85 of the respective families.
For the prototype-ROBIN the PPC405CR was selected for the following reasons:
•
Sufficient safety margin with quoted performance of 370 MIPS @ 266 MHz
•
Good match of features for application
•
Compatibility with integrated processor of successor86 FPGA families.
As seen from the measurement results, this processor was just powerful enough to fulfil the latest87
requirements with 2 ROLs in the GE environment. As the required processing power scales pretty
linearly with the number of channels, an adequate processor would need approx 50% more
performance. For the final ROBIN a similar processor was selected: the PPC440GP, operating at 466
MHz. The 440 core is an out-of-order dual-issue machine with two execution pipelines combined with
two integer units and one load/store unit. The 405 core contains a single-issue execution engine, and
although it is a scalar processor, the core can perform loads and stores in parallel with arithmetic logic
unit (ALU) operations. Reasons to select this type of processor for the final ROBIN were:
82
83
84
85
86
87
•
Expected performance improvement by factor of 2 – 2.5, quoted value 932 MIPS @ 466 MHz
•
Memory size and bandwidth improvements by factor of 2
•
Cache sizes increased to 32 kB each for instruction and data (factor 2 and 4 respectively)
•
Identical bus interface to prototype-ROBIN
See also UK-ROB project documentation at http://www.hep.ucl.ac.uk/atlas/rob-in/processor.html
The PPC4xx family has moved to AMCC in the meantime.
IBM PPC405CR, Intel 80200
Like PPC405GPr or IOP321
XILINX Virtex-2Pro and Virtex-4 are available with integrated PPC405 CPU cores.
When the prototype-ROBIN was developed a maximum readout-ratio of 10% was assumed. For the final
ROBIN this number was increased to 20%. The prototype-ROBIN conforms to this higher figure as well.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
63
Chapter 5 - ROBIN
•
Software compatible to prototype-ROBIN
This performance gain expected from the quoted MIPS values was confirmed by the results available
from “the Embedded Microprocessor Benchmark Consortium”88, which showed a performance ratio
between 2 and 2.5 for different application types (Figure 31).
Figure 31: PowerPC performance comparison
A comparison published by IBM89 led to similar results. A simple test application derived from the
ROBIN buffer management code run on a sample set of 20k fragment pages produced 472 kHz
processing rate on the PPC405CR compared to 1054 kHz on the PPC440 mezzanine (Figure 32),
hence a factor of 2, as expected.
Memory subsystem
The memory subsystem of the CPU is made up from two 512MBit DDR-1 memory chips, providing
128MB total size and 1GB/s maximum bandwidth. The memory is used to hold the application
program and a copy of the operating system90.
Despite the lower frequency compared to the high-speed serial signals of the ROL/S-Link interfaces,
the DDR memory is the most challenging part of the PCB layout. To avoid complications, the ROBIN
88 Source no longer available on the internet.
89 http://www-03.ibm.com/chips/power/powerpc/newsletter/jun2003/lead.html compares a PPC440GP with a
PPC405GPr (already 50% higher clock speed than the PPC405CR)
90 The ROBIN uses only a simple boot-loader and monitor program as operating system. Refer to software
section.
64
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
follows the implementation guidelines91 for DDR-modules of the memory vendor.
Figure 32: PPC440GP mezzanine
DDR-memories transfer data with both edges of the clock signals, but address and control signal
change on the rising edge only. Signals can be logically divided into the three groups CLOCK (C),
ADDRESS/CONTORL (A) and DATA (D), as shown in Figure 33.
The clock – which is a 120 Ω differential signal – from the memory controller inside the CPU is
duplicated in a PLL clock buffer and distributed to the 2 memory chips. The address and control
signals are referenced to the clock only, while data signals are masked and captured by separate
DATA-MASK and DATA-STROBE signals, on a per-byte basis. The “A” group and the “D” group –
all are 50 Ω single-ended signals – have particular timing requirements, which are translated into
length-matching requirements for the traces on the three segments “CPU to series resistor”(38 – 72
Figure 33: DDR memory subsystem
91 http://download.micron.com/pdf/technotes/TN4607.pdf
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
65
Chapter 5 - ROBIN
mm), “resistor to memory” (10 – 15 mm) and “memory to terminating resistor” (5 – 15 mm). The total
length must be between 60 and 80 mm. Length matching of group “A” has to be adjusted to +/- 2.5mm
with respect to group “C”. Length matching of group “D” has to be adjusted to +/- 2.5mm with respect
to the data-strobe signal of the corresponding byte group and to +/- 12mm with respect to group “C”.
Values of R-s and R-p are 22 Ω and 33 Ω respectively.
Board-level simulation was not performed prior to PCB production, as the ROBIN implementation
with only 2 discrete memory devices has a significantly lower capacitive load than a memory-module
design92, which operates with 8 to 20 devices in parallel. A brief cross-check of some signals after the
production run shows acceptable signal behaviour, see Figure 34.
Figure 34: DDR memory simulation
92 http://download.micron.com/pdf/misc/ddrregrev1.2.pdf
66
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
FLASH memory
An 8 MB FLASH-memory is connected to the external bus of the CPU, which loads its boot-code and
ROM-monitor after power-up from that FLASH. As the external bus is also connected to the FPGA,
factory programming of the FLASH can be performed from the PCI interface, via a special FPGA
design, while the CPU is held in reset. A special one-time-programmable sector93 is available in the
FLASH which is used to store a serial-number and production data.
External bus
The PPC440GP has two bus interfaces, a PCI-X bus and a so-called “external” bus (Xbus). As the
PCI-X protocol is very complex the simpler Xbus is used to connect the CPU and the FPGA. The
Xbus operates at 66 MHz with 32 bit non-multiplexed address and data and includes support for
external masters and DMA.
The example in Figure 35 shows a series of write cycles. A burst is started by asserting a chip-select
signal (PerCS), together with address (PerAddr), data (PerData) and read/write control (PerWE,
PerOE). A ready-signal (PerReady) from the peripheral controls the insertion of wait-states after each
data word. The burst is terminated with PerBlast.
Figure 35: PPC Xbus device-paced burst write timing
93 This feature is called „secure silicon“ (SecSi) sector. It provides 256 bytes of user data. This sector can be
mapped to the area of sector 0 with a special command.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
67
Chapter 5 - ROBIN
5.3.1.5 PCI Interface
The maximum data volume to be transferred from a ROBIN to the host PC is around 100MB/s for
three channels, at 20% request-rate. Compared to this, the bandwidth required for messages passing is
very low: 20 byte per fragment request, 420 byte per delete request of a group of 100 events, which
sums up to less than 1MB/s per channel.
The PCI interface is copied from MPRACE-1 and uses a commercial PCI-to-local-bus device from
PLX, the PCI965694. On the PCI side it provides a 64bit/66MHz bus, compliant to PCI specification
2.2. On the local side – towards the FPGA of the ROBIN – there is a 32bit/66MHz multiplexed bus
(Lbus). The throughput of the device is clearly limited by the local bus to 264MB/s, but the safety
margin95 is sufficiently high: a factor of 2.5. In addition to simple direct-slave PCI mode, the chip
provides 2 scatter-gather DMA channels and local-bus-to-PCI direct-master96 access. The PCI9656
supports PCI-interrupts via mailbox-registers and via a user-pin from the FPGA.
The Lbus interface of the PLX device operates in “J”-mode and is very similar to the PCI bus. A
typical direct-slave write operation is displayed in Figure 36. The PXL device accepts the command
from PCI, arbitrates for the Lbus using LHOLD and LHOLDA, places the address on the multiplexed
address/data bus LAD together with the address strobe ADS. In the next cycle the first 32-bit word is
placed on LAD. When READY is signalled from the FPGA the second 32-bit word is sent, together
Figure 36: PLX local bus timing
94 http://plxtech.com/download/PCI9000/9656/databook/PCI_9656BA_Data_Book_v1.3_13Jan09.pdf
95 264 MB/s available vs. 100 MB/s required.
96 Both DMA and local-master modes are “initiator” modes on PCI. In DMA mode, the PLX device generates
the addresses and access sequence. In local-master mode the device on the local bus generates addresses and
access sequence.
68
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
with the terminating BLAST signal. Master transfers from the FPGA are done in the same fashion, but
the FPGA arbitrates then for the LBUS, and the direction of the signals is reversed. A 256 byte deep
FIFO in the PLX decouples the two buses on direct-master writes and provides for maximum
throughput.
During normal operation the only communication mechanism is message-passing: direct-slave writecycles to send messages to the ROBIN and local-master write-cycles to send corresponding responses
from the ROBIN to the host. This scheme allows a very efficient utilisation of the PCI-bus, especially
when multiple PCI devices are operating concurrently.
5.3.1.6 Network Interface
A GE interface employs functional blocks at the physical and the link layer, which are frequently
implemented using two discrete devices, a PHY and a MAC. The selection of the PHY was driven by
the aim to support both electrical and optical media. The MARVELL 88E1011S can be used for both
GE media and the major MAC interface protocols – GMII and TBI – are supported as well. The
prototype-ROBIN implemented both optical and electrical interfaces with automatic media detection.
On the final ROBIN only the electrical media was retained.
For the MAC functionality three candidates have been analysed: a XILINX core97, the LSI810498 and
the Intel IXF100299. The LSI 8104 doesn’t support the GMII interface which makes it less flexible
when attaching a PHY device. Also there is no support for VLAN tagging. The XILINX core was
considered to be an interesting alternative in principle, but not mature for the prototype-ROBIN. The
IXF1002 had all required features and was selected for the prototype-ROBIN. On the final ROBIN,
the XILINX core was used, which saved some PCB area but more important allowed easy
enhancement with optimisations like MAC address filtering, IP alignment (see 5.3.2.3 ) and input
queue.
Due to the characteristics100 of the network traffic in ATLAS TDAQ a buffer for a large number of
incoming messages had to be implemented. For this purpose a single ZBT device101 of 2MB size is
attached to the FPGA and operated in dual-port emulation mode. Packets received by the MAC core
are stored into this external device and a corresponding packet descriptor is pushed into a FIFO. The
descriptor is read by the CPU, which subsequently receives the packet from the ZBT. The bandwidth
of the ZBT is 264 MB/s, which allows for concurrent READ and WRITE at full GE line-speed.
5.3.1.7 Auxiliary components
Besides the main components described in the previous sections, some auxiliary functions had to be
implemented. The most prominent ones are being briefly described here:
97 http://www.xilinx.com/products/ipcenter/GMAC.htm
98 http://www.lsi.com/DistributionSystem/AssetDocument/files/docs/techdocs/networking/8101_8104_DS.pdf.
This device is obsolete.
99 Device is obsolete, no online document available.
100On rare occasions burst containing up to 1000 message may be broadcast, using a non-reliable network
protocol (UDP). A large packet buffer helps to minimise packet loss even in such cases.
101 Cypress CY7C1371D-100AXC,
http://download.cypress.com.edgesuite.net/design_resources/datasheets/contents/cy7c1371d_8.pdf
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
69
Chapter 5 - ROBIN
•
Power supply
The main supply voltage is 3.3V, taken from the PCI connector. This voltage drives the
buffers, GE PHY, PCI interface, the ROL/S-Link interfaces and auxiliary logic. Also, all other
voltages are generated from this source. Switching regulators are used to generate 1.5V for the
FPGA core supply, 2.5V for the GE PHY, PCI9656, CPU and DDR memory subsystem, 1.8V
for the CPU core supply and 1.25V for the termination of the DDR memories. For testing
purposes and factory programming, the power can be supplied through an ATX-style 8-pin
connector. The power consumption of a ROBIN is around 15W under normal operating
conditions.
•
Control functionality
Board level control functionality is realised in a XILINX XC2C256 Coolrunner-2 PLD
“CPLD”. A board-level RESET signal is derived from signals from a voltage-monitor device,
a manual switch and the PCI reset. An 8-position DIP-switch is used to control configuration
parameters, e.g. related to JTAG and RESET. The system-clock buffer is supplied with a
frequency derived from a 66MHz oscillator, which is simply fed through for normal operation
or divided for testing purposes. During power-up the CPU reads configuration values 102 from
an I2C-controller implemented in the CPLD. LEDs are driven to indicate RESET and FPGAconfiguration status. The topology103 of the board-level JTAG chain is also defined by the
CPLD.
•
JTAG interface
Figure 37: JTAG routing
Not all components104 of the ROBIN support JTAG, hence using JTAG for board-level testing
is limited. Instead, JTAG is mainly used for configuration and debugging. The CPU has a
102 So-called „strapping“ values.
103 For testing purposes FPGA, CPU, PLX and PHY devices can be arbitrarily connected to a JTAG chain.
104 E.g. all memory and TLK2501 devices and don’t support JTAG
70
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
private JTAG connector which allows connecting a JTAG-debugger105 for software download
and testing. Also, the FLASH memory can be initialised that way. JTAG access to the main
JTAG chain – comprising Control-CPLD and FPGA – is possible from 3 different JTAG
control ports: JTAG connector, PLX PCI-bridge and CPU (see Figure 37). Multiplexing
between the three sources is done in a small XILINX XC9536XL PLD “JPLD”, the content of
this device is assumed to be stable106. The JPLD is factory-programmed via a separate JTAG
port on the same JTAG connector. The content of the CPLD is stable but can be upgraded from
the JTAG connector107 or PLX JTAG port. FPGA configuration is normally done at power-up
time from the CPU JTAG port, but can be done as well from the JTAG connector108 or the PLX
JTAG port109.
•
Testing is supported by a number of on-board test-points for voltages and clock-signals, a 50
signal mezzanine connector attached to unused FPGA pins and CPU based serial and fastethernet ports.
5.3.1.8 Realisation
The assembled PCB of the final ROBIN is shown in Figure 38. On the lower left side there are 3
optical transceiver sockets110 with the TLK2501 SerDes devices next to them to the right. Above the
Figure 38: Final ROBIN
105 Abatron BDI2000 or BDI1000 debuggers are used.
106 The prototype-ROBIN does not have the separate JPLD, instead all functionality is in the single control
CPLD. This makes in-situ firmware upgrades of the CPLD impossible.
107 The JTAG connector is used for factory-programming of the CPLD content.
108 The JTAG connector is normally used to verify proper JTAG operation of the FPGA during factory testing
109 The PLX JTAG port is normally used during factory-programming and to test new FPGA bit-streams before
making them resident.
110The optical transceivers are hot-pluggable and not installed on this picture.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
71
Chapter 5 - ROBIN
socket there are the GE connector (number “2” written on) and the GE PHY device. The ATLAS logos
covers the free space towards the FPGA, which was occupied on the prototype ROBIN by the GE
MAC device, which is now internal to the FPGA. Above the logo there is one of the buffer memory
chips (there are 2 chips per channel, one each on opposite sides of the board) and below there is the
switching power regulator which generates 2.5V and 1.5V from the 3.3V source on the PCI connector.
Right to the logo one can see the FPGA and right to the power regulator the PCI bridge. Next to the
bridge there are the PowerPC processor, the power regulators for 1.8V and 1.25V and one of the two
chips of the processor memory. At the edge of the PCB are the (smaller) connector for the serial port
of the processor and the (larger) connector for the processor's JTAG interface. Right to the FPGA we
have two more buffer memories (top) and the network packet buffer. At the top right corner of the
PCB there are two more JTAG connectors, the reset-button, DIP-switch and the CPLD. The RJ45
connector of the private Ethernet interface of the processor is located in the middle on the right side of
the board. On the bottom side of the PCB there are mainly passive components like resistors and
capacitors and the remaining memory devices. The connector in the top most left corner is used to
power the card in stand-alone operation, for example during factory testing.
5.3.2
VHDL Firmware
The FPGA design is composed from a number of VHDL modules, jointly developed at the three
institutes Mannheim, NIKHEF and RHUL and additionally including IP-cores from CERN and
XILINX. Figure 39 displays all major VHDL modules.
The main elements used to implement the receiving and storing of event data comprise a ROL-Slice,
which is replicated 3 times. The DMA engines on the path towards Lbus 111 and Ethernet incorporate
multiplexers, to select one channel at a time. The CPU reads data from various sources, including the
UPFs, via a multiplexer. The Control/Status block represents all functionality not related to the main
data flow, e.g. CPU access to control and status registers which are available in each of the other
modules, reset signals, writing test patterns etc.
In total, the FPGA handles 7 different clock domains. One of them can be viewed as the “system”clock running at 66MHz. Modules crossing domains from system clock to one of the I/O clocks are
indicated with bold borders. Each HOLA-module receives an individual input clock from the
TLK2501 transceiver. Together the HOLA-modules share one transmit clock. Both HOLA input and
output clocks operate at 100MHz. All buffer memories and the MAC transmit path share a common
125MHz clock. Finally there is a 125MHz receive clock at the MAC. CPU, Lbus-interface and MAC
dual-port-memory operate at the system clock. A reasonable grouping of the modules leads to the
following functional blocks:
•
Input channels: the FPGA internal part of the ROL Slices
•
Request channels: the Lbus and MAC dual-port-memory related blocks and the Ethernet MAC
(shared with output)
•
Output channels: the two DMA engines with the associated header and buffer FIFOs, plus the
111 The FPGA doesn’t connect directly to the PCI bus, but through a commercial PCI bridge device. The local
bus (Lbus) of that device is a 32 bit multiplexed bus.
72
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
Ethernet MAC (shared with message)
•
CPU access: although almost invisible in the diagram there is complex code related to address
decoding and accessing control and status registers.
Figure 39: ROBIN VHDL modules
5.3.2.1 Input channels
Each ROL-Slice connects to an external ROL-Interface, which consists of a TLK2501 SerDes device,
which in turn connects to an optical transceiver. The serial data received over the ROL is converted to
16-bit parallel data by the SerDes and passed on to the HOLA Core (see 5.3.1.2 ). Data to be sent back
along the ROL (via S-Link Return Lines) is serialised by the SerDes before being transmitted over the
ROL. S-Link is a unidirectional data path with flow-control and low-speed return lines, which are not
used in the ROBIN application. The data-path is 33 bit wide – 32 bit of data plus 1 control bit, valid
data is indicated via a single flag. If flow-control is turned on at the receiving end, the sending side
will stop writing data after some delay, typically a few cycles plus the delay of the media (e.g. the
optical fibre). The HOLA core already provides some amount of buffering to compensate the media
and source delay. Additional buffering and handling of the upper layers of the transmission protocol is
done by the S-Link handler. This module performs a number of consistency checks on the incoming
packets and indicates errors to the subsequent stage:
•
Framing: data section must be encapsulated by one begin-of-frame (BOF) and one end-offrame (EOF) control word. If required, the S-Link handler terminates a erroneously framed
packet by inserting an EOF control-word.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
73
Chapter 5 - ROBIN
•
Packet length: the minimal length of the data section must comprise the start of the header up
to the L1ID. In addition, the actual number of words (header + payload + trailer) must match
the packet length indicated in the corresponding trailer field.
•
Transmission errors: flagged by the HOLA-core upon CRC mismatch
•
Format: the format field in the header must match the expected format.
Apart from the HOLA-core data can be fed into the S-Link handler by a data-generator or a test-input
FIFO. The data generator generates S-Link packets of programmable size at maximum speed and is
used for stand-alone performance measurements. The L1ID-field of subsequent fragments is
automatically being incremented and the pay-load section is filled with a test-pattern. In contrast the
test-input FIFO is controlled directly by the CPU and is therefore slow, but more flexible than the
data-generator and is used to test all possible S-Link error conditions.
All words going into the S-Link handler are forwarded to the buffer memory module, certain words
are accompanied by special flags (e.g. error flags, control-word-flag, L1ID-flag, run-number, etc). The
header-positions which trigger the L1ID and run-number flags are programmable.
Buffer-management controls the storing of incoming fragments into the buffer memory and the lookup of the fragments on requests from the TDAQ system. It is the central function of the ROBIN. The
approach is based on the paged buffer management scheme developed by the UK group [UKROB],
with some optimisations applied. All available buffer memory is logically arranged into pages of fixed
but programmable size (typically 2kB). Every fragment occupies at least one of these pages. A freepage-FIFO (FPF) and a used-page-FIFO (UPF) decouple the buffer-manager module from the
corresponding buffer-management code on the CPU.
The buffer-manager module (Figure 40) retrieves free pages from a free-page-FIFO one by one. It uses
flow-control to stop data if there are no more free pages available. The free-page information
generates the starting address for the fragment and this address, followed by the data, is sent to the
buffer-input FIFO112. A bit in the buffer-manager control register selects if the BOF/EOF control words
are retained or stripped off the data stream. For every processed memory page one information record
(4 words) is written to the UPF, which contains the following information:
•
Page status: error code, fragment termination
•
Page address, length of data, trigger-type field
•
L1ID field
•
Run number field
The information is collected during the transmission of the page and transferred to the UPF in a single
cycle of 128 bit. On the CPU side, the UPF represents 4 distinct 32 bit registers and is updated (read)
when reading the top-most one (run-number).
The size of the decoupling FIFOs is 1k entries for the FPF and 512 entries for UPF.
112 The buffer-input FIFO provides a FULL flag which could also be used to trigger flow-control. However, the
memory bandwidth is guarantied by design to always exceed the maximum bandwidth of the link.
74
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
The buffer-input FIFO queues the data to be sent to the Buffer Memory and the addresses to which
they are to be written.
The final stage of each input channel in the FPGA is the buffer memory controller which arbitrates
between write request (from buffer manager) and read requests (from DMA). The relative priority of
input and output is compile-time programmable to accommodate different requirements. However,
with the buffer memories operating at 125MHz, the available bandwidth is in the order of 400MB/s
thus allowing a simple 1:1 ratio without imposing any limitation. On the reading side there is a tripleport interface, which accepts requests from the two DMA engines and from the CPU. The latter is used
only for debugging and must not be activated during normal operation, as the CPU timing is
incompatible with the input timing requirements. A read-request is just a command with a starting
address, which flushes the buffer-output FIFO and initiates data transfer from the memory. Transfer
pauses when the FIFO is full. A “stop” command ends the transfer, flushes the FIFO and prepares the
controller for a new command.
Figure 40: FPGA buffer manager
5.3.2.2 Request channels
The TDAQ system can send requests to the ROBIN through the host PC or directly via the network
interface and both paths can be active simultaneously. At the VHDL level, the two interfaces are
handled similarly: a request is stored into a dual-ported memory and a request descriptor is written to a
FIFO. For the network case, this functionality is explained in [section output channels]. For PCI, the
implementation is even simpler: requests from the PCI host arrive via the Lbus at two different address
areas, one corresponding to an internal DPR, the other to the descriptor FIFO. The host software
calculates addresses, sizes, and keeps track of the filling state of the request buffers; hence, the FIFO
and DPR have smaller sizes than the equivalent units do on the network interface, and there are no
special controlling units needed.
5.3.2.3 Output channels
The output channels are used to send fragments from the buffer memories as well as any other
message data towards the downstream TDAQ system. In all regular cases, a DMA engine controlled
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
75
Chapter 5 - ROBIN
by the CPU accomplishes the data transfers. There are a few exceptions to this rule, which allow the
Lbus host to bypass the messaging mechanism and directly read particular information from the
ROBIN:
•
FPGA design id: a 32 bit value identifying version, revision and author of the code
•
S-Link real-time status: link-up and x-off
•
Embedded serial port: utility module, which connects a host terminal program to the ROBIN
processor.
The two DMA engines for Lbus and network access are almost identical, but certain discrepancies in
the control information prevented to use a generic DMA engine for both channels.
Every DMA engine monitors its header-FIFO and first retrieves a DMA descriptor block, which
defines the number of header words to follow the DMA descriptor, the number of data words from the
buffer memory, the buffer memory channel and the starting address in the memory. The DMA engines
subsequently transfers all words from the header FIFO. Next, it issues a read-command to the buffermemory-controller and transfers the requested amount of data from the buffer output FIFO. Zerolength header or buffer fields are possible. For example, a response to a non-data request will have a
zero-length buffer field. Similarly, a multi-page fragment going to Lbus will not need header data for
subsequent pages.
For Lbus, two additional words – the PCI destination address – are required in the DMA descriptor.
When the DMA engine has data available, the Lbus interface stores some words in a small (32 words)
internal FIFO, arbitrates for the local bus and subsequently sends the data using the direct-master
mechanism of the PLX PCI bridge (see section 5.3.1.5 ). The same mechanism – but a different
address range – is used to write into internal registers of the PLX device.
The MAC DMA-engine provides a feature which simplifies generation of IP-packets. The IP protocol
uses 16-bit alignment while the FPGA uses 32-bit alignment. One of the descriptor bits enable the
CPU to supply modified IP packets with a 16 bit padding field right after the IP header, which makes
the payload section 32 bit aligned. The DMA engine removes the padding element before sending the
packet to the MAC unit and sets a flag for the MAC transmitter. Two other flags indicate the start and
end of packet words.
Figure 41 displays the MAC unit, which is composed of three major sub-units. The MAC transmitter
accepts data and flags from the DMA-engine and stores packets in a double-buffered memory – one
packet per buffer – which also converts the 32 bit wide data path to the 8 bit data path of the MAC
core running at 125 MHz. The MAC core is a commercial113 IP-core from XILINX, which supports –
together with the external PHY device – a 1000Base-T Gigabit-Ethernet link, including standard
Ethernet statistics. Configuration of the MAC core is done via a set of control and status registers,
accessible from the CPU. The MAC address is shadowed in a separate register and used by the MAC
receiver to validate uni-cast packets. The receiver converts the 8 bit data stream from the MAC core
into 32 bit wide packet data, performs IP-realignment by inserting a 16 bit padding element. Packet
113 XILINX supported the research activity by donating this IP-core.
76
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
data are stored in the external MAC input buffer114 and the associated packet-descriptors into a
descriptor FIFO. Each packet descriptor indicates the length (in words) and the starting location of its
associated packet in the memory. The depth of the FIFO is configurable to 511, 1023, 2047 or 4095
entries. A value of 1023 allows buffering of 1k maximum sized Ethernet packets with the present 2MB
external memory. Flow-control is turned-on when either the buffer or the FIFO is almost full.
Figure 41: FPGA network interface
5.3.2.4 CPU access
The CPU communicates with the FPGA via a 32-bit non-multiplexed big-endian bus. The Lbus in
contrast is a little-endian bus and to avoid confusion the CPU bus is converted to little-endian format
right at the edge of the FPGA.
The CPU internally generates a number of chip-select signals, with programmable timing. The FPGA
uses two of them to distinguish different areas. “Write”-timing is single-wait-state fixed for both areas.
“Read” timing is as follows:
•
Internal registers, FIFO and the buffer memories (for debugging): variable timing, minimum
two wait-states, non-cacheable.
•
Dual-ported memories: one wait-state fixed timing, cacheable.
The dual-port area must be placed in a separate area to take advantage from burst-mode access of the
CPU bus, which is only available when caching is enabled. Clearly, caching cannot be enabled for any
kind of FIFO access. The total number of registers and FIFOs feeding the read-data-multiplexer is 43
(plus the 2 DPRs), which requires a prioritised multiplexing scheme, which puts the DPR areas at the
higher level and all registers at the lower level with a slower access.
114 The external memory is not a true dual-ported memory device, but a ZBT memory. The dual-port memory
controller in the FPGA emulates dual-port functionality by alternating read and writes access on a cycle-bycycle basis. The guaranteed input bandwidth is still above Gigabit-Ethernet line speed (available: 4 bytes @
33MHz, required: 1 byte @ 125 MHz = 4 bytes @ 31.25 MHz).
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
77
Chapter 5 - ROBIN
The CPU interface has to generate a number of read-enable and write-enable signals to load or unload
the various FIFOs.
There are three DMA channels available, which can be used instead of the normal read-mode access to
the three UPFs. CPU software can select between DMA mode and regular mode.
5.3.2.5 Resource utilisation
The design is implemented into a XILINX XC2V2000-5FF896C FPGA and consumes more than 2/3
of the available resources115, as shown in Table 6. About 70 timing-constraints116 guide the timingdriven mapping process, which distributes the synthesised logic across the FPGA fabric. Floorplanning is not used, as most of the modules are self-contained and automatically placed close to the
corresponding I/O pads and/or according to the constraints.
Utilisation Summay
Logic Utilization
Used
Total Number Slice Registers
14,279
Number used as Flip Flops
14,265
Number used as Latches
14
Number of 4 input LUTs
13,357
Logic Distribution
Number of occupied Slices
10,555
Total Number 4 input LUTs
15,569
Number used as logic
13,357
Number used as a route-through
1,227
Number used for Dual Port RAMs
634
Number used as Shift registers
351
Number of bonded IOBs
589
IOB Flip Flops
706
IOB Dual-Data Rate Flops
1
Number of Block RAMs117
50
Number of GCLKs
8
Number of RPM macros
24
Total equivalent gate count for design
3,602,489
Additional JTAG gate count for IOBs
28,272
Available
21,504
Utilization
66%
21,504
62%
10,752
21,504
98%
72%
624
94%
56
16
89%
50%
Table 6: FPGA resource utilisation
The source-code is implemented in 120 VHDL files with 43000 lines (including comments), plus 18
IP-cores. According to the different tools used at the various institutes the VHDL is quite portable and
can be synthesised with XILINX XST, Mentor Graphics Precision and Synplicity Synplify. A VHDL
design needs five compilation steps to generate the executable, which is frequently called a “bitstream”. The total run-time for this project is about one hour. Synthesis and mapping 118 take in the
115 The figure „98% of slices occupied“ doesn’t mean that the device is actually full, as there are still LUTs and
Flip-Flops available. The logic is just distributed across the entire device and a denser packing will allow to
put more functionality into the device. However, placement and routing will become more difficult.
116 Definitions of clock frequencies, setup-to-clock, clock-to-output and invalid-paths.
117 The present implementation uses only a single buffer on the MAC transmit path, which saves one BRAM
compared to the estimation.
118 The mapping process is timing-driven and generates a placement. Place-and-route doesn’t do another
placement, just the routing, which makes is quick.
78
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
order of 25 minutes each, translate, place-and-route and bitgen use 5 minutes or less each.
5.3.3
Software
The software running on the ROBIN CPU consists of two parts: an open-source boot-loader/rommonitor called “u-boot” and the “robin” application software. Both packages reside in the on-board
flash-memory. After power-up, the CPU loads its initial configuration from an I2C-attached memory
and subsequently loads the boot-code from the flash memory. U-boot then initialises the internal
peripherals like DDR-memory controller, serial port, external bus and MMU and copies itself into the
main memory (DDR). MMU initialisation is done simply by mapping each memory region119 to a
separate MMU entry. The MMU has space for 64 TLB entries, so there is still space for additional
mappings. After being initialised, the two serial ports provide terminal access to the monitor
functionality. U-boot listens on both serial ports for incoming data and switches the active port to the
one most recently active. This is useful when no serial cable is attached and serial communication has
to run via the FPGA.
After initialisation, u-boot loads the application binary from a pre-defined address, or enters terminal
mode if no application is present. Terminal mode provides a number of useful commands to inspect or
modify environment variables, registers and memory, to load programs, and for rudimentary
debugging. More information on the commands is given in [ROBMAN].
U-boot on the ROBIN does not support a graphical debugger like “gdb”, so the only debugging aid are
“printf” statements, which can be turned on or off under user control if the application was compiled
with the debug-option. Furthermore the download of a new binary over the serial interface takes a
couples of minutes. Both properties of the system make the software development process a bit
inconvenient. To enable shorter turn-around times and improved debugging the main ROBIN
application can be compiled in “host”-mode and as such run on the development machine (a Windows
PC). In this case a large part of the FPGA functionality is replaced by a simple software model and the
PCI communication by a shared-memory structure. A corresponding test program is able to attach to
the shared-memory structure, alternative to the PCI communication. This way, both host and
embedded application can be executed under the control of a graphical debugger on the development
machine, which is much more convenient. Apart from the entire PCI-based message handling the
emulation supports the functional test of the fragment processing via upload of event fragments and
interrupts, basically everything except network operation. For the test program emulation and physical
ROBIN are equivalent, apart from the different initialisation sequence.
The ROBIN application program is a monolithic, single-threaded executable, composed from 10 “.c”
source-files and 20 “.h” header-files, in total120 11500 lines of “C”-code and 6500 comment lines. The
10 modules implement different functional groups as follows:
•
robin.c: the “main” module, which steers the execution of the program with the taks-loop
•
robin_bufman.c: handles UPF and FPF including error handling, manages fragment look-up
119 Pre-defined memory regions are: DDR-memory, internal peripherals, every chip-select area on the external
bus
120 Lines counted with the „cccc“ utility, available at the open-source repository http://www.sourceforge.net
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
79
Chapter 5 - ROBIN
and deletion
•
robin_pci.c: receives and converts requests from the PCI host, initialises Lbus DMA-engine
•
robin_net.c: receives and converts requests from the network, initialises network DMA-engine
•
robin_msg.c: final decoding and dispatching of requests
•
robin_init.c: start-up initialisation and run-time re-initialisation
•
robin_util.c: utilitiy functions like terminal output, profiling, etc.
•
robin_serctl.c: JTAG configuration of FPGA
•
robin_bist.c: self-test functions
•
uboot_stubs.c: interface to u-boot system calls
The first five modules comprise the core functionality and are performance critical. The remaining
modules contain code which is used infrequently.
The following section explains the most important functional elements of the modules, which are:
•
Main task loop
•
Buffer management and garbage collection
•
Request - response messaging scheme
•
Instrumentation
•
Configuration and operational monitoring
•
Initialisation
5.3.3.1 Main task loop
The tasks comprising the main task loop are buffer management and request handing. Figure 42 shows
a simplified view of the loop. The GetFragment function reads the status of the UPF and processes the
available pages, up to a programmable maximum number. This is done for all channels in sequence.
Accepted pages are stored in the page database. The next task in the loop is to check for an incoming
request from the PCI host. If this fails121, the loop checks for a request from the network. A request
from any of the interfaces is converted into a generic format – however with interface specific
information attached – and sent to the request dispatcher. The dispatcher calls the appropriate
command – in the example either a data request or a delete request – which finally acknowledges the
request via the appropriate DMA-engine.
A few parameters steer the priority of the tasks in the loop: the maximum number of pages processed
by GetFragment and a pre-scaling value for each of the tasks. The default values are 10 for the
maximum number of pages and 1 for pre-scaling, which provide good performance in the standard (=
low request-rate) situation. When a high read-out ratio is required (e.g. ALL fragments have to be
121 In the environment of the ROBIN, only one of the two interfaces will issue requests at a high rate.
Therefore, no prioritisation is done here.
80
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
retrieved), the GetFragment pre-scaler must be increased to allow more time for request handling.
Figure 42: Software task loop
The actual implementation of the task loop is slightly more complex, e.g. the update of the FPF is
done in a separate low-frequency task and not directly attached to the Delete function. There is an
“idle”-task performing some internal bookkeeping. Input from the serial port branches to a simple
debug interface, inside the application.
5.3.3.2 Buffer management
Every page has to be analysed for a number of conditions: new fragment, fragment termination,
transmission errors, format errors and sequence errors. The acceptance policy of the ROBIN is to
reject as few fragments as possible. Only if the fragment does not contain a L1ID field, it is rejected 122.
For all other error conditions, the fragment is kept and flagged as suspicious. Status and length
information from all pages of a fragment are recorded in the corresponding fields of the first page.
Figure 43: Software buffer manager
Figure 43 displays the data structures used in the buffer management scheme. The UPF provides the
122 A limited number of rejected pages is stored in a separate buffer, which can be retrieved by a special
command.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
81
Chapter 5 - ROBIN
basic information on a per-page basis, which is converted into a FragInfo structure. This structure is
then copied to the associated field in the MgmtEntry element, which is indexed by the page number.
Next, a hash-key is created from a configurable number of low-order bits of the L1ID field, typically123
16. The corresponding entry in the hash-list points to the first element (head) with this key in the item
list.
The MgmtEntry pointers are used to create a double-linked list of pages with the same hash-key,
terminated with a link to page 0, which is reserved. New entries are added at the head of the hash-lists.
Figure 44 shows the mechanism for a few single-page and multi-page fragments.
Figure 44: Buffer manager database
Deleted fragments are removed page by page from the linked list and the pages are pushed onto a
stack, which keeps all available free pages. When appropriate, the FPF is re-filled from this stack.
ItemList, hashTable and freePageStack are static vectors and consume in the order of 10MB for all
channels, which can easily be accommodated by the system.
The interaction between the FPGA part and the software part of the buffer management is shown in
Figure 45.
123 Possible values fort he number of bits are 10 to 20. 16 bits will lead to exactly one entry per hash-list with
64k pages and linear ordering of the fragment. More bits will produce sparse occupancy, less bits will
increase search time.
82
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
Figure 45: Fragment handling
5.3.3.3 Request – Response messaging
All communication between the TDAQ-system and the ROBIN uses a message-based 124 request –
response scheme. The generic format of a ROBIN message is very compact and this format is used
directly for bus-based transactions. In general, a request message specifies a command, selects a
channel, indicates the address to which a reply shall be sent and contains information specific to the
command. After completion, the ROBIN sends a response message to the source 125 indicated by the
request. A transaction-identifier helps to verify proper handshaking. Receiving requests from a PCI
host is simple, as it stores them into the Lbus DPR already in the right format.
For transactions over the network, the ROBIN has to accept requests defined by TDAQ messagepassing protocol [DC022] – which are transported via UDP/IP – and to convert them into the generic
format. Due to limitations of the early TDAQ software, a single ROBIN initially had to present itself
as multiple ROSes, each with a single data channel126. The current implementation assembles the data
from multiple channels, just like a ROS-PC does. The UPD implementation of the ROBIN supports
response messages of arbitrary size, but only single packet requests. This does not impose a limitation,
as data requests and delete requests smoothly fit into a single Ethernet packet and the TDAQ-system
does not send other messages to a ROBIN over the network. There is one exception to the request –
response rule: network delete messages are typically sent to a multi-cast address; hence, the return
address is unknown and no response is generated.
In both cases, the media specific software module (robin_pci or robin_net) attaches a media specific
control block to the (converted) request. For PCI communication, the control block keeps only the
return address. For network communication, the control block must be initialised with proper
Ethernet-, IP- and UDP header information. Subsequently the message with the control block is
delivered to the request dispatcher (see also Figure 42).
124 A message in the context of this document is a logical unit of information, transporting control and/or data.
125 Normally the response goes back to the originator of the request, but this is not required by the protocol.
However, the present implementation uses always the same media for request and response.
126 Channel selection is done via different IP-addresses on the same MAC-address.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
83
Chapter 5 - ROBIN
Once the ROBIN has completed the command, it generates a response message. Data responses
require a TDAQ-wide defined “ROB”-header [RAWFMT] preceding the fragment data from the
buffer. All other responses use a generic ROBIN header, eventually followed by command dependent
data127. All responses except data responses are handled with two unique DMA transfers:
•
Main transfer: DMA-descriptor, header and response data are assembled into a contiguous
memory block and pushed128 into the Lbus header-FIFO. The DMA-engine sends the block to
the host PC.
•
Completion: A special block with only a single magic word is send via DMA to the starting
address for the response, indicating end of transmission.
Data responses require special handling, as they may transfer data from different, non-contiguous,
memory pages in a buffer. For PCI, the first transfer takes the header and the first page of the
fragment. Subsequent pages are sent without header data. A completion transfer terminates the
response. The mechanism for request and response is shown in Figure 46.
For the network, there are two additional complications:
•
The maximum Ethernet packet size requires larger pages to be split
•
Ethernet- and IP-header are needed on all packets and must be constantly updated
Figure 46: PCI fragment request and response
The software overhead (compared to PCI) to process network responses is significant, even if a
completion transfer is not required.
127 For example, a delete request generates a list of L1IDs, which did not match any available fragment.
128 Memory-to-memory DMA might be used in this case, but is not yet implemented.
84
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
5.3.3.4 Instrumentation
The use of the GNU profiler utility “gprof” requires a file-system to deposit the profiling data, which
is not available on the ROBIN. In order to get parameters though for the TDAQ modelling approach,
all critical functions are equipped with custom timing measurement macros. Up to 16k execution-time
values per interesting code section are recorded. The average values are printed on the terminal at the
end of the application. The performance penalty compared to the non-instrumented code is in the order
of 10%. As shown in Figure 47 CPU performance is mainly consumed by three functional blocks:
handling of fragments, handling of messages and updating of the free-page-FIFO. The execution times
obtained from running the instrumented code were use to predict the performance of the final code. A
comparison of the predicted performance and the real performance will be presented in section 6.1 .
An important debugging tool is printout to the serial terminal. Clearly, printing information
permanently has a disastrous impact on the performance and makes the application virtually unusable.
To overcome this, a debug print macro allows switching the printout dynamically on or off. For
example during testing of the error handling, the applications starts-up with debug output disabled,
then printout is enabled just before erroneous fragments are inserted, to display all details of the
fragment processing. The application – with debug output disabled – still reaches about 50% of the
performance of the non-instrumented version.
Figure 47: Software execution times from profiling
The timing and debugging macros do not generate any code in the regular (non-instrumented)
application version.
5.3.3.5 Configuration and initialisation
A set of about 40 configuration parameters (section 8.2 ) controls certain aspects of the the operational
behaviour of the ROBIN. The initialisation sequence scans the FLASH memory for defined
parameters during the start-up of the ROBIN application and loads either the defined values or the
hard-coded default values.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
85
Chapter 5 - ROBIN
Approximately one third of these parameters relate to debugging or to special testing conditions. For
example, the input handler module of the ROBIN can be forced to start without waiting for a valid
input fragment, which is useful when new hardware is attached to the input links. Test functions enable
running the ROBIN without an external data-source but with an internal data-generator instead.
Fragment size and various data patterns can be specified in this mode of operation. A verbose mode
can be enabled when the the debug version of the application software is loaded, which prints detailed
information of every activity to the serial port.
Configuration parameters relevant for regular operation are for example IP address, detector source
ID, maximum size for input fragments – above which truncation occurs – and start/stop signalisation
for the input handler module.
Most of the configuration parameters can be arbitrarily modified at run-time by the host software.
Some of the parameters – like the size of the memory pages – need a re-initialisation of the channelspecific management structures. This can be done while the other channels remain operational. A few
parameters however, for example the IP address, require a full re-start of the ROBIN application. This
is accomplished by writing the new value of the parameter to the FLASH memory, followed by a reset
of the ROBIN. The new value is then loaded from FLASH during the initialisation sequence.
A consistent view at the configuration parameters for both the local ROBIN application and the host
software is achieved by generating a special header file “robin_cfg_map.h” from a version of the
ROBIN application, which is compiled and executed on the host. This file contains the textual
identifiers, default values and classification flags which are needed for proper handling of the
configuration parameters by the host software.
5.3.3.6 Operational monitoring
The ROBIN performs operational monitoring by gathering statistics of all important performance
figures at the input and the message interfaces, in total about 60 individual numbers (section 8.3 )
including:
•
Numbers of consumed, deleted and available pages
•
Numbers of received, requested and deleted fragments
•
Numbers of data-request, delete and other messages
•
Error counters
In addition, the ROBIN monitors the input link status, the temperature alarm signal and the most
recent event ID and builds histograms of the fragment sizes (16 bins in units of ¼ of the page size) and
buffer filling state (16 bins).
The accumulated values are reset upon a special command from the host, which also has to compute
performance figures, if desired, as the ROBIN does not use reference timing information.
5.3.3.7 Self-test functions
The Built In Self Test (BIST) is defined as a short and simple standalone facility to test the
86
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
functionality and connectivity of the main components of the ROBIN and is part of the ROBIN
application. The BIST is executed directly after the low-level initialisation of the CPU and
configuration of the FPGA every time the application starts including after each reset. An extended
version, EBIST, can be run if a more extensive buffer test is required.
No external devices or cables are required to initiate the test; the ROBIN need only be connected to
power via PCI or a power cable. The results can be seen on a serial terminal and are provided within
the status response of the ROBIN. If the application is able to start and run this self-test, the PPC
processor and its RAM can be assumed to be functioning.
The tests of the functional components are divided into functions that are explained below. The FPGA
test is the first test. Then the network test is performed once. The order of the other tests, which are
executed for each channel, is not critical because there are independent. However, it is probably most
efficient to leave the EBIST extended buffer test to last, as it takes the most time to execute and the
results of the other tests functions can be read during this time.
FPGA Test
The design ID of the FPGA firmware is located in a register of the FPGA and shows the version and
the revision of the firmware loaded. The expected version and revision is defined in the robin.h file.
The version must be the same although the revision is permitted to be higher than expected. In the
latter case normal operation should be possible but some new or corrected functionality in the
firmware might be unavailable.
An incorrect version or a revision lower than expected generates a critical DesIdMissing or
DesIdWrong error message respectively. Because the BIST cannot reasonably run under this
condition, the application stops with an error. A higher revision results in a warning only.
If the firmware version is acceptable a reset of the FPGA registers is performed and the buffer status
registers are read. If the buffer initialisation failed then a ResetError flag is set.
Buffer Test
The buffers store the fragment data for each ROL and each buffer has to be tested separately. Their
capacity is 64 MByte each. Each buffer has a buffer controller implemented in the FPGA and, in
addition, is connected to the external bus of the PPC.
There are three types of buffer tests:
•
Data test – tests data lines (32)
•
Address test – tests address lines (24)
•
Extended buffer test – write marching “1” and “0” respectively and verify read data
The address and data tests are quick and integrated in the BIST; the extended buffer test takes ~10 sec
per buffer and runs in the EBIST only. This is the only difference between the BIST and the EBIST.
At the beginning of each buffer test, the chosen buffer is enabled with the enableBuffer(rolId)
function. If this fails, a BufEnableError is set. In the data test one bit is shifted left (starting with
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
87
Chapter 5 - ROBIN
0x00000001) and written to the first address, one junk word to another address and then the value of
the first address is read out. If the read out value is not equal to the data the test causes a BufDataError.
For data bits lower than 24 the corresponding bit in the address test should also return an error,
because these data bits are used there, too. The address test writes the address as data to the address in
the buffer for every 24-bit address. So e.g. the address 0x200 contains the value 0x200. If the read out
value is not equal to the address the test causes a BufAddrError. One address line is tested at a time. If
there is no BufDataError, a connection failure of this particular address line can be assumed. If many
address and data errors occur, it is probably a hardware error on the clock signals to the buffer.
The extensive buffer test in the EBIST writes a marching “1” to every address in the buffer. This is
like the data test, but covering the whole buffer address space at a randomly selected starting value.
After filling the buffer (~2 sec), the data is read out and compared (~8 sec).
The steps are repeated for a marching “0”. If the values read out are different to the marching “1” and
“0” respectively, a BufExtError is caused. This is not a critical error, but a warning is issued.
Network Test
The network interface is realized with a MAC (implemented in the FPGA on the ROBIN) and a PHY
component. The PHY, which can be controlled via the MAC, has a loop-back function that is used in
this test.
If the function used to initialize the MAC function in the FPGA firmware returns an error the
NetMacCfgError is set.
The interface between the MAC and the PHY is GMII. If the GMII does not receive a ready signal
after sending a command, a timeout occurs and the test causes a NetGmiiTimeout.
The PHY is set into loopback mode (if this fails NetPhyCfgError is set), and a packet of data is written
to the transmit buffer of the MAC. The receiver buffer is read and the values are compared with the
original data. If they are not equal, this results in a NetRxError. The receiver buffer is read again and if
these values are different to the written data, an NetTxError is caused.
A RxError with no TxError indicates that the data has been routed through the PHY back to the FPGA
correctly, but could be read out only at a second attempt.
A NetRxError and NetTxError together show that the data are not written correctly or the read values
are inconsistent with the written data. This could indicate defective data lines between the FPGA and
the PHY.
The loopback functionality of the PHY is turned off at the end of this test.
ROL Test
The connection to the Read-Out-Links is routed via the FPGA to a TLK 2501 SerDes for each ROL. In
the FPGA a bypass is implemented which allows the direct control of the TLK control lines. The built
in Pseudo Random Bit Stream (PRBS) test and the loop-back function are used for testing the
connectivity of all data lines to the TLK and its functionality.
First the FPGA internal resets are asserted to initialize the FPGA (clearing FIFOs, etc.). The bypass
88
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
and the PRBS are then enabled. If the Status register does not show the bypass enabled, a
RolEnableError is set.
Next the RxFifo is read to remove any leftover data and to free the FIFO for the PRBS results. The
PRBS is stopped and the results are read from the RxFifo. The TLK indicates a successful PRBS test
with the RXERR bit set. If not, the test returns a RolTlkPrbs error. The RxFifo is again read to provide
a pre-defined status.
The internal loop-back function of the TLK is used to compare sent and received data. For this random
data values are created from the lower time register. Inconsistent data causes a RolTlkLoop error.
If the RolExtLoop environment variable is set to 1, an optional external loop-back test follows. For
this test, a fibre must be connected from the output of a ROL to the input of the same ROL. If this test
is not successful, a RolTlkExtLoop warning is issued.
The bypass functionality of the TLK is turned off at the end of this test.
5.4 Installation and commissioning
The installation procedure is defined by assembling the ROS-PCs with the ROBIN cards and the PCIe
NIC, mounting of the PC in the ROS rack and attaching power and network cables and the S-Link
fibres. Also, the operating system is installed. The subsequent commissioning involves the integration
of the ROS-PC with the detectors and the TDAQ software framework. In the course of this procedure
the functionality of the system and the correctness of the interconnects between all subsystems are
verified. To manage the large number of nodes a graphical database utility was created, the ATLAS
RackWizard, see Figure 48. The main pane (top left) shows a single level of ATLAS building USA15,
where every box represents a rack. The view of one particular rack is displayed on the top right pane,
Figure 48: Rack wizard
with 8 ROS-PCs in the lower area and the switches at the top. The lower right pane shows the rear
view of one of the PCs, with the installed components. Finally the lower left pane displays the entry
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
89
Chapter 5 - ROBIN
for one of the ROBINs with the serial number and “geographical” position within the rackWizard's
address space. The full database information can be accessed from there via the link to the MTF129
database (Figure 49).
Figure 49: MTF Entry
At the start of the LHC on September 10th 2008 the ATLAS TDAQ system was commissioned with the
full dataflow system and an HLT farm with 6800 processing cores130 for L2 and EF corresponding to
approx. 35% of the final size [ATLCOM].
5.5 Summary
The requirements on the ROBIN as a component of the ATLAS ROS have been presented. A single
PCI card must accept 3 optical input channels operating at 160MB/s. Event fragments arrive at up to
100kHz on each link and are stored in 64MB page-organised buffers. Requests to transmit or delete
event fragments arrive over a PCI or GE interface at up to 21kHz, depending on the ROS architecture
– bus based or switch based. In addition to these basic functions complex operations related to
configuration and monitoring are required. Continuous updates and adaptation to various detectors
require a flexible design of the component. The hardware platform of the ROBIN is composed from a
high-density FPGA device, which handles all high-speed and high-rate real-time operations and data
movements. Data transfer towards PCI or GE are performed by DMA engines inside the FPGA, which
provide direct paths between buffer memories and target interfaces. The FPGA application consists
mostly of custom VHDL code plus external library elements for the S-Link and GE interfaces.
A high-performance embedded processor performs all complex task like message processing, event
129 MTF is the “manufacture and test folder “of the main ATLAS ORACLE database.
130 Many installed machines are equipped with multi-core processors, so the number of machines is lower.
90
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 5 - ROBIN
bookkeeping and monitoring. The processor interacts with the FPGA through a 32 bit local bus with
DMA capabilities. The embedded software is a monolithic “C”-application which runs on top of a
simple boot-loader and monitor program. All binaries are resident in a local FLASH memory and
booted after power-up.
Sophisticated tools for self-testing and to aid software development are available. A CERN-based
component inventory database keeps track of all ROBIN cards and provides status and location
information.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
91
Chapter 5 - ROBIN
92
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
6
Results
This chapter presents the results obtained with the ROBIN, presented in different views relating to the
different requirements areas. The first section describes the test setups and measurements used to
characterise the performance of the ROBIN as a component. The system performance of the base-line
ROS is presented thereafter. A reasonable test environment for the switch-based ROS is still under
preparation, therefore the system performance of the switch-based ROS will be presented elsewhere.
Issues which occurred during the development process are presented in the last section, which also
covers reliability aspects.
6.1 Stand-alone
The basic setup used to assess the performance of a ROBIN requires:
•
a ROS-PC with the ROBIN installed in one of the PCI slots
•
a PC emulating the detector via a multi-channel S-Link source card (FILAR131)
•
a PC emulating the TDAQ system
•
a GE switch, connecting the ROS-PC, the ROBIN and the TDAQ emulator
In reality, the detector emulator is frequently substituted by the internal data-generator of the ROBIN,
which has almost identical properties. The TDAQ emulator needs to represent different elements of
the TDAQ system: a DFM, which is responsible to delete events, plus L2PUs and SFIs to request
events. Again, there are simplifications: to measure the performance in the PCI environment no
external TDAQ emulator is used. Instead, all requests and deletes are generated by the ROS-PC as
otherwise the increased load on the ROS-PC would limit the performance. In the network environment
the performance of a single TDAQ emulator machine may not be able to saturate a ROBIN. Also, the
requests generated by L2PU and SFI are equivalent to the ROBIN. Therefore, the typical setup in this
case will contain one DFM emulator and two SFI emulators. The same configuration can also be used
to verify that ROBIN and ROS-PC are functionally equivalent132 from the network point of view. The
GE switch must properly handle flow-control messages and should provide virtual output queues (see
chapter 3.2 ) to minimise packet loss.
As described in section 5.3.2 , the movement of data in the ROBIN is handled by hardware.
Concerning the incoming bandwidth over the S-Links and the processing rate of memory pages there
are basically no limits. The nominal 160MB/s of the S-Link are exceeded by far by the internal data
generator (up to 250MB/s) and by the external data source (up to 200MB/s). The maximum fragment
rate of the data generator had to be limited to 200kHz in order to stay approximately within the
specifications even for small fragments. On the outbound interface, data is sent by the DMA engine
from the buffer memory to the PCI bridge or to the network port. The buffer memory has a total
bandwidth in the order of 400MB/s, hence the DMA can provide more than 200MB/s to the output
port while the input is running at nominal speed concurrently. The throughput of the output ports is
limited by the internal design of the PLX PCI bridge to 264MB/s or to the line-speed of 125MB/s in
131 FILAR documentation is avialable at http://hsi.web.cern.ch/hsi/s-link/devices/filar/
132 The ROBIN must respond in the same way to the same messages.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
93
Chapter 6 - Results
the case of the GE interface respectively.
The expectation for the stand-alone performance is therefore, that it is under regular operating
conditions completely dominated by the performance of the on-board processor and that the impact of
the hardware design is only visible – in the form of bandwidth limitation – for large fragments or very
high rates.
A common complication in assessing the ROBIN performance with a stand-alone setup is the missing
synchronisation between event generators and requesters. In the real system the L1 issues event
identification to the HLT system and thus assures that requests address only existing events. The standalone setups however have to use free-running event generators. Tracking the event number in this
scenario can become complicate, in particular if multiple nodes – e.g. L2, SFI and DFM - need to be
synchronised to the locally generated events. The synchronisation mechanism has to monitor the value
of the most recent event number processed by the ROBIN, which is transmitted via the status of every
response message. Another field of the status indicates two cases of missing events: “lost” or
“pending”. A fragment is “lost” when it is not available in the buffer and the event number of the
incoming fragments is (logically) larger than the number of the requested fragment. This is typically
caused by a delete message which was issued asynchronously. In contrast, a fragment is “pending”
when the request addresses an event with a number larger than the most recent one received by the
ROBIN. In this case the request should be repeated.
6.1.1
PCI access
Performance measurements were done by exercising the ROBIN via the PCI message-passing
interface with the program testRobin. Initially, it generates a number of requests for each of the
ROBIN channels and submits them into the hardware queue of the ROBIN, which allows for a
maximum of 10 requests133. For every response coming back a new request is submitted to keep the
pipeline busy. After the steady-state of the operation is reached the program measures the execution
time for a certain number of events, typically in the order of a million, to exclude caching effects and
the influence of the various hardware FIFOs. The results of measurements [ROBJINST] with an
external detector emulator are shown in Figure 50. Performance is expressed as sustained L1 rate as a
function of the request rate. For small fragment sizes (100 and 200 words) the L1 rate drops linear
with the request rate, which conforms exactly to the expectations. The characteristic parameters in this
region are the processing time per fragment tF and the processing time per request tR. TF includes the
times to register a new fragment in the bookkeeping system and to delete it later on. TR includes the
times to process the request message and to setup the corresponding data transfer. From the diagram
we can derive the parameters134 to be tF = 2.04ns and tR = 4.6ns. The actual data transfer is an
overlapping operation, executed by the DMA controller in the FPGA, hence the impact of the
fragment size becomes visible only when the transfer 135 time is equal or greater than the processing
time, which occurs around 1kB. For larger fragments tR therefore is a function of the fragment size
133 Recent tests with an increased hardware queue size of 64 entries (~20 per channel) didn't show a significant
improvement.
134 Note that the rates are shown per channel, the total rates are 3 times larger.
135 The maximum output bandwidth towards PCI is around 220MB/s
94
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
with tR = size[kB]*4.6ns.
Figure 50: Actual PCI stand-alone performance
The requirements on the ROBIN are indicated by the two arrows, where the left one corresponds to the
TDR assumptions for a bus-based ROS under full load and the right one represents the more recent
understanding, in both cases for a nominal fragment size of 1kB. There is a significant safety margin
available even in the extreme case. In addition to the performance calculation based on the parameters
obtained from the measurements the performance can be estimated using the timing values obtained
from the instrumented code. The fragment processing time is composed of the times to read the
fragment from the FPGA and to enter the fragment into the bookkeeping structure plus the times to
receive and process the delete message and finally the time to push the freed page back to the FPF. In
total, this sums up to 2.02µs which is very close to the 2.04µs of the initial calculation. The request
processing time is composed from message reception and processing time, summing to 3.7µs. The
latter value is smaller than the one from the calculation above, which indicates sub-optimal conditions
during the measurements. The calculated values are shown in Figure 51, where the “INSTR” prefix
indicates values from the instrumented code.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
95
Chapter 6 - Results
Figure 51: Calculated PCI performance
6.1.2
Network access
In case of the network setup only preliminary results are available, obtained with the simplified
program dcRequester instead of the combination of DFM and SFI/L2PU nodes and an initial version
of the network communication protocol. The same set of measurements was executed and the ROBIN
parameters obtained are tF = 1.85ns and tR = 12.5ns. The improvement in tF is a result of the more
efficient fragment deletion procedure in this arrangement, where a single message deletes the
fragments from all three channels and no acknowledge needs to be returned. In contrast, PCI delete
messages are sent individually to the channels and expect an explicit acknowledge. The increase in t R
is due to the much more complex mechanism required to handle the UPD data request messages as
compared to the PCI message interface. Figure 52 Shows the performance of the regular application
(“normal”) for fragment sizes of 900 and 2000 bytes. In addition, the performance of the instrumented
code (“timing”, lower performance) and the code without monitoring function (“xfast”, higher
performance) is displayed. For comparison, the estimations based on the parameters from the
instrumented code (tF = 2.13ns, tR = 12.4ns, label “INSTR”) and from the function fit (tF = 1.85ns, tR =
12.5ns, label “CALC”) are shown. It can be seen that already at a fragment size of 900 bytes the
performance drops below the prediction for high request rates, which is probably caused by the
increasing load on the machine running the requester program. For larger fragments the bandwidth
limit136 of approx. 75MB/s is approached for all 3 application versions (“timing”, “normal” and
136 The FPGA design used for these measurements required a non-overlapping copy of the network packets first
from the buffer into a FIFO, then from the FIFO to the network which limits the max. bandwidth to
96
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
“xfast”), which corresponds to 12.5kHz request rate for 2kB fragments.
Figure 52: ROBIN network performance
The influence of the operational monitoring can be estimated by comparing the XFAST and
NORMAL curves. XFAST yields in the order of 10kHz higher a L1 rate at the same request rate.
6.1.3
Event bookkeeping
As explained in section 5.3.3.2 the bookkeeping procedure uses some of the lower order bits of the
L1ID as hash-key to index the events. The number of bits used has an influence on the processing time
of the regular event lookup (hashing) during adding or removing of a fragment as well as on the
special functions required to remove orphaned fragments from the buffer (garbage collection). The
maximum number of bits which can be used is limited to 22 by the memory required to store the hashtables (corresponding to 24MB for three channels).
All test setups used so far generate and delete events sequentially, such that the distribution of
fragments in the ROBIN buffers is very regular and compact. Under realistic operating conditions
however this will be different, for example there will be events which require processing times in the
EF much longer than the nominal buffering latency of the ROBIN. Also, the rate of lost network
packets might be non-zero, leading to residual fragments. The optimisation of the number of hash-bits
and also the bit positions composing the hash-key can later on be done on the running system, based
(1/(1/200MB/s + 1/125MB/s)) = 77MB/s.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
97
Chapter 6 - Results
on the following results.
6.1.3.1 Hashing
Without experience from the experiment a regular distribution of L1IDs in the buffer is assumed. For
this approach, the best performance of the hashing is achieved by using as many bits as possible, with
the upper boundary being the maximum number of events in the buffer, which is defined by the
number of buffer pages. With the typical value of 32k pages the target number of bits for the hash-key
Figure 53: Effect of hash-key size
is 15. Figure 53 displays the two software functions which are affected by the hashing mechanism fragment processing and event deletion – and the impact of the size of the hash key for a buffer size of
32k pages.
6.1.3.2 Garbage collection
Garbage collection consists of 2 steps, a linear scan of all buffer scan for fragments, followed by the
selective deletion of fragments which fall outside the validity range. As shown in Figure 54, the
deletion time is almost constant for a small number of buffer pages (“DEL-4k”) while for a larger
number of pages the dependency on the size of the hash-key is obvious (“DEL-32k”). The building of
the fragment list is equal in both cases (“BF-32k”, “BF-4k”) for hash-key sizes up to 12 bit, which
corresponds to the number of pages of the small buffer. If the hash-key size is increased further in this
case, a negative effect is introduced by the sparse occupancy of the hash-table, which is caused by the
regular event processing in the test setup.
98
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
Figure 54: Garbage collection performance
6.2 System performance
System performance has been evaluated first at the level of a standard ROS-PC. This can be viewed as
an extended test setup. In addition, initial measurements of the performance of the entire dataflow
system are available from the first beam period.
6.2.1
Bus-based ROS
The baseline bus-based ROS is evaluated using an arrangement of a ROS-PC – equipped with 4
ROBIN cards – attached to 2 PCs emulating the dataflow system via 2 GE links (Figure 55). Data
input is taken from the hardware data generators inside the ROBINs.
Figure 55: ROS test setup
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
99
Chapter 6 - Results
The requirements have evolved since the TDR, which requests 3kHz of EB plus 17kHz of L2 for 2
ROLs at 100kHz1 L1 rate in case of the “hot-spot” ROS. The current assumptions used for the latest
measurements [TIPP09] are 3kHz of EB plus 18kHz of L2 for 3 ROLs at 100kHz L1 rate. In the even
more demanding case of a scan of the full sub-detector by L2 the RoI size grows to 12 ROLs which is
equivalent to an EB rate of 21 kHz. Figure 56 shows the dependency of the total request rate
(composed of a fixed EB portion and a variable L2 portion) to a ROS-PC on the fragment size, at a
fixed 100kHz L1 rate, for the standard configuration and the recently optimised configuration. It can
be seen that the original performance requirement is met by a standard ROS-PC, however only for
small fragments at the RoI size of 3. After optimisations of the software, in particular by using a uniprocessor kernel, tuning off hyper-threading and the security features of Linux, and by tuning the
interrupt-coalescence of the network interface card, the situation improved and the performance for
RoI sizes of 3 is now well above the requirements. It is to mention that the average load on the
individual ROBINs in the system is relatively low, per channel only 6kHz for an RoI size of 2 and
8kHz for an RoI size of 3 respectively at the target figure of 21kHz total request rate on the ROS-PC.
High EB rates are problematic still, because the output bandwidth limit of the 2 GE links is reached
already at 20kHz for the nominal 1kB fragments.
Figure 56: Performance of the bus-based ROS
6.2.2
ATLAS Dataflow
According to [ATLRDY] the full installed ATLAS dataflow system consists of 1500 computing nodes,
which is a large fraction of the final size of about 2300 nodes. From the computing nodes 850 were
100
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
quad-core machines with dual network interfaces, allowing to use them as both L2 and EF processing.
On average, each node runs between 4 and 5 applications concurrently. For the final event building a
target bandwidth of 5GB/s – generated by 59 EB applications – is envisaged. The full system was used
in two configurations, one with data preloaded into the ROS-PC and running with emulated L1
triggers. Here, different L2 trigger menus are tested. The other configuration was the full operation
during the first beam period between August and November 2008.
In the test with emulated input a L1-rate of 60kHz and an EB bandwidth of 3GB/s could be achieved
for realistic trigger menus. The load on the ROS-PC varied between sub-detector – as expected – an
was in the range of 5 to 30kHz, as shown in Figure 57.
Figure 57: ROS request rate, emulated data
The analysis of the first beam period is not complete yet, however an initial result is constituted by the
successful operation of the entire system during the full period, with the recording of 1PB of data.
6.3 Issues
During the design of the ROBIN hardware a few issues required re-layout or re-work. As shown in
section 5.3.1.4 the layout of the DDR memory signal is critical. However, careful layout is not only
required for the primary signals (address, control and data) but also for the power supply of the
termination resistors. Improper decoupling and too long trances lead to spurious memory errors on the
first version of the PCB.
After completion of the volume production frequent errors on the optical links were observed, after
connection to the external data sources. These errors however did not occur during the optical loopback tests, where data is transmitted locally on a ROBIN. The reason was an error during the
manufacturing process of the 100MHz crystal oscillator providing the reference frequency for the
TLK2501 SerDes devices. The result of this error was the locking of the oscillator frequency to the
wrong frequency of the crystal after power up. This phenomenon is known as spurious mode
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
101
Chapter 6 - Results
oscillation137 (Figure 58).
The measured frequency derivations were in the range of 500 to 1500ppm, with the specification being
+/- 50ppm. The oscillator manufacturer had to supply new components, which were mounted to the
boards after delivery to CERN.
Figure 58: Spurious mode oscillation
A third issue came up with memory errors observed on the network packet buffer, occurring however
only on the UK batch of the cards. The debugging process took several week and the reason was found
only by accident – an incorrect assignment in the VHDL code generated a bus-contention between
FPGA and memory during the addressing cycle and in turn a ground bounce effect on the address
lines. A minor difference in the UK and German PCBs – power layer thickness of 18µm vs. 35µm –
made the German cards tolerating the ground bounce while the UK cards didn't. The last PCB-related
problem which concerns the readout of the FPGA die-temperature via the remote temperature sensor
MAX1618. This device is connected to the base-emitter diode of an on-chip NPN transistor in the
FPGA. The P-N junction resistance is temperature and current dependent and the resistance ratio for
two different currents is a measure for the temperature. Unfortunately, a few high-speed traces of the
memory system cross the two analogue signals between the diode and the sensor and the introduced
noise prevents reliable measurements. The error was only detected after the production of the preseries ROBINs, after the layout was in principle completed. The only way to avoid another re-layout
was to remove the traces to the FPGA in the Gerber data-set and to replace the buffering capacitor
with a diode. Although the diode is not on-chip but only in the proximity of the FPGA and its
characteristics is different from the base-emitter diode of a transistor, this work-around at least allows
to estimate the chip temperature. As visible from Figure 59 the transistor sensor follows the actual
temperature – measured via a PT100 sensor – quite well, while the diode sensor has a significant,
variable offset. With software-calibration the accuracy can be tuned to approximately 10°C, which is
sufficient for the application.
An unexpected property of the external bus of the processor – which connects to the FPGA – poses a
general performance limitation on the design. According to the datasheet, the external bus can provide
a bandwidth of up to 266MB/s at 66MHz bus frequency. As explained above, the main use case for the
communication between processor and FPGA is read and write access to FIFOs inside the FPGA,
137 Spurious mode information: “AT Crystal Spurious Modes”, http://www.ecliptek.com/tech/spurmodes.html
102
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
typically performed as a series of accesses to a fixed address. In this configuration the effective
transfer speed from a FIFO to the main processor memory is only in the order of 20MB/s, because the
external bus controller does not activate its burst mode. Improvements could be made by modifying
the FPGA to implement dual-ported memory areas for the message data from the PCI and network
interface. These memory areas are mapped into the processor's memory as cacheable regions, which
allows the external bus controller to use bust-mode read access. The transfer bandwidth was improved
to 55MB/s.
Temperature characteristics
Temperature readout [°C]
140
120
100
Transistor
Diode
Temperatur
80
60
40
20
0
Time
Figure 59: Temperature sensors
With a measured write bandwidth of approximately 300MB/s into main memory these numbers
correspond to net read bandwidths of 22MB/s and 67MB/s respectively. For the reverse direction138 the
integrated DMA controller can be used to perform a memory-to-memory transfer into the command
FIFOs. While this allows to overlap the transfer time with other processing, the total improvement is
only marginal due to the additional setup required for the DMA. Finally, an attempt was made to use
the DMA controller to copy the UPF information in larger chunks from the FIFOs into the main
memory. Unfortunately there were spurious errors in the received UPF records which could never be
completely resolved. Therefore the DMA feature is not in use for this purpose.
6.4 Reliability
The ROBIN cards are fully commissioned and in operation since more than a year and some
experience concerning the reliability is available. In the first year of operation a relatively high number
(approximately 50 occurrences) of hardware failures were observed. The majority of these failures
could be repaired by rewriting the content of the on-board flash. There is no clear explanation for this
behaviour, however after a modification of the firmware upgrade procedure the number of incidents
went down significantly. So this issue is believed to be a minor software problem. From the remaining
failures about 10 exhibited similar errors related to incorrect operation of the PCI interface and all of
these card were from the UK production batch. During the error analysis the PLX PCI interface was
138 The single-cycle write bandwidth is only marginally higher: 28MB/s
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
103
Chapter 6 - Results
removed on a few of these board for further investigation. The inspection of the solder pads indicated
that a problem with the PCB is very likely, which is known as “black pad” phenomenon139. As a result
certain solder pads are not properly or not all wetted by the solder. As the device is mechanically fixed
onto the PCB by the correctly soldered pads an electrical connection can nevertheless exist but is not
reliable. Frequently the devices pass the factory test but fail later on in the field, as it was the case with
the ROBIN cards.
Figure 60: PLX unwetted pads
In Figure 60 it can be seen, that a few pins at the lower left corner look very different after the removal
of the PLX device. A precise analysis of the situation would require a destructive analysis of the
affected boards, which was not performed. About 20 cards exhibited errors which might be related to
this “PLX problem”. Fortunately the error rate went down to a single defective board in 2008. Overall,
the initial failure rate of the ROBINs was relatively high, but has now reached a plateau at about the
same level as the ROS-PC and its other components which is below 1% per year. The high initial
failure rate and the drop to a plateau is commonly described as “bathtub curve”140.
Figure 61: Bathtub curve
As the name indicates, one has to expect a rising error rate after some time of operation. To date, the
system is still working at low rates of failures.
139 See e.g. http://circuitsassembly.com/cms/images/stories/pdf/0301/0301raytheon.pdf
140 See e.g. http://www.weibull.com/hotwire/issue21/hottopics21.htm
104
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 6 - Results
6.5 Summary
The results for the ROBIN components and for the ROS system have been presented. In stand-alone
PCI operation the ROBIN can handle up to 27kHz request rate at 100kHz event rate, which is well
above the requirements. The corresponding operation over the network yields only 15kHz, however
there is still room for optimisations. A few issues were encountered after the volume production was
finished, related to quality of layout, components and PCB. However, none of the issues was serious
and the ROBINs are working very reliable since installation.
The baseline bus-based ROS meets the performance requirements for small event and RoI-sizes,
which is acceptable for most of the ROS-PCs. A few ones only will be exposed to higher rates and can
be tuned by reducing the number of active channels. The entire installed ATLAS dataflow system
consisting of more than 1500 machines has shown stable operation and good performance during the
first beam period in autumn 2008.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
105
Chapter 6 - Results
106
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 7 - Discussion and Conclusions
7
Discussion and Conclusions
The final chapter shows that the initial goals with respect to performance and functionality have been
met and presents prospects for future developments based on the present implementation.
7.1 Performance Assessment
The results presented in chapter 6 verify that the performance goals of the ATLAS TDAQ system can
be met. At the component level, the ROBIN card performs superior to the requirements in the baseline
bus-based architecture. The network performance has already been improved since the time of the
measurements presented here, by adding an alternating buffer mechanism to the network output path
and by software optimisations. Latest results141 indicate that the 21kHz request rate requirement can be
met. The design paradigm to use an FPGA for high-speed and high-bandwidth functions and a
standard microprocessor for the more complex but slower functions was a success. The FPGA
firmware is mature and requires modifications only on relatively long time-scales (order of 6 months
or more), for example after modification of the input data format. Most of the modifications during the
regular maintenance and release procedures can be implemented by software updates, which is very
advantageous due to the much more convenient development process compared to FPGA firmware
(see also section 4.3 ). In the end, much more operational monitoring and error checking could be
implemented on the ROBIN as initially planned. The firmware (FPGA) and software package are well
integrated into the TDAQ software release framework, such that the maintenance of this custom
component is well organised.
At the level of the ROS-PC two bottlenecks were identified: processor performance and memory
bandwidth. Recent tests with a new type of PC motherboard using a dual-core processor and a faster
memory subsystem show a significant performance boost, such that a ROS-PC with 12 active channels
is only limited with respect to request rate by the performance of the ROBIN cards and with respect to
output bandwidth by the number of GE links. The TDAQ system as a whole has been operated in test
runs and during the first beam period with very good results and the required performance and
reliability for the operation during the first few years seems to be available. Final results however can
only be obtained when the LHC machine provides particle collisions at design luminosity, which is
expected to start in autumn 2009. The quality of the individual sub-detectors and the physics
algorithms will then define the actual rates the TDAQ system has to sustain.
To address performance requirements which cannot be fulfilled with the standard bus-based ROS, the
following scenarios are envisaged:
•
With the lower beam luminosity of the initial period, higher performance will be required on a
few ROS-PCs of the LAr detector only. It is foreseen to lower the number of active channels
on these PCs until acceptable performance is achieved. Most probably they will be equipped
with two ROBIN cards.
•
With increasing beam luminosity the number of affected ROS-PC might rise. At this point, the
ROBINs of the relevant sub-detector will be operated in switch-based or hybrid mode. The
141 The results from the improved networking functionality will be presented elsewhere.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
107
Chapter 7 - Discussion and Conclusions
current view of the hybrid mode is to direct L2 requests directly to the ROBINs via the
network interface, while EB requests and the delete commands will still be passed via the
ROS-PC.
•
Optimisations of the TDAQ software will eliminate the current duplication of requests for
event building.
Overall, the ATLAS TDAQ system is ready to operate at design performance for the first LHC period
up to 2012.
7.2 Prospects
Despite all these good results it is clear from sections 5.3.2 and 6.1 that the safety margins at the
ROBIN are quite small - FPGA resource utilisation is very high and the processor performance is
already at the limit, at least for network based operation. Adding new functionality or supporting even
higher rates will be close to impossible. Also, PCI as an ubiquitous host interface is fading out and
being replaced PCIe. Finally, the upgrade scenarios for LHC – phase 1 and phase 2 – will require some
modifications to the TDAQ architecture.
7.2.1
Short term developments
To address the PCI/PCIe issue and to increase the component performance a variation of the ROBIN is
currently under development, which supports a PCIe interface, a faster processor and a larger buffer.
The general design and all142 remaining components are copied from the existing ROBIN. The
modifications to FPGA firmware and software will be minimal, such that a common set of source files
can be used. The new PCIe interface is of single-lane type and supports the same bandwidth as the
present PCI interface. The new processor is the pin-compatible device PPC440GX, which operates at
667MHz instead of 466MHz and provides an on-chip 2nd level cache, which the present processor does
not have. The buffer memory is increased by a factor of 2, which enables to provide 64k buffer pages
of 2k each guaranteeing single page fragments for all possible sizes at 100kHz L1 rate. The
expectation for the performance of the PCIe ROBIN is that a request rate of 25 to 30kHz in network
mode can be sustained, corresponding to 75 to 90MB/s bandwidth for 1kB fragments.
The PCIe ROBIN will be produced in a quantity of approximately 100 cards. During the regular
maintenance process failing ROS-PCs or the ones at the end of their life will be replaced by PCIe
capable machines, equipped with PCIe ROBINs. This will be done primarily for sub-detectors with
high performance requirements. If PCI-based PCs will still be available, the current ROBINs can be
re-used for the other sub-detectors.
In the area of the switch-based ROS certain improvements have already been made and will be
documented elsewhere. In particular the single outbound packet buffer which posed a bandwidth limit
on the tests presented in 6.1.2 was replaced by a dual-buffer mechanism, which allows to reach GE
line-speed for large fragments. The network protocol was simplified and the ROBIN now aggregates
data from all 3 channels into a single UDP message. Further optimisations are under investigation, for
142 Apart from the substitution of obsolete parts with recent versions.
108
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 7 - Discussion and Conclusions
example changing the Ethernet MTU to 4kB, which should bring the request rate for standard 1kB
fragments to above 20kHz and the bandwidths close to GE line-speed already at 2kB fragments.
Finally, tests with recent motherboards supporting dual-core processors and providing better memory
bandwidth have been made, using the present ROBIN in bus-based mode. The results [ROSRT09]
indicate that the performance of a new ROS-PC is about twice that of the present standard ROS-PC,
using the existing ROBINs.
The switch-based ROS and the gradual replacement of ROS-PC with more recent machines – with
present or new ROBINs – in critical areas allow for significant performance improvements at the
system level during the next few years of ATLAS operation.
7.2.2
Upgrade Phase-1
The LHC phase-1 upgrade is scheduled for the year 2012 and will include modification of the pixel
sub-detector via an insertable B-layer (IBL) and in general higher detector occupancies due to higher
beam luminosity. No significant changes will be applied to the detectors in general and to the readout
electronics. The HLT system will be improved by a fast track trigger (FTK) which will operate on the
inner detector data and provide tracking information to the L2PUs with a latency of 1ms. The dataflow
system will be exposed to the same rates as today, but to event sizes larger by 50%. For the ROBINs
and the ROS, this will not pose significant problems, as they are limited rather by rate than by
bandwidth.
However, an interesting project is foreseen concerning the IBL development, which allows to
prototype a ROB-on-ROD module, installed in parallel to the standard S-Link output from the pixel
ROD. The goal hereby is to implement the ROBIN functionality for a single channel on a small
mezzanine card, minimising real-estate and power consumption. The interface towards the ROD will
be the S-Link connector. On the TDAQ side there will be a GE interface – supporting both electrical
and optical media – for the dataflow and a second electrical GE interface for control functionality. To
address the issue of the decoupling of the two subsystems TDAQ and detector/ROD the mezzanine
will support power-over-Ethernet (PoE143), driven via the control interface. To minimise the load on
the driving switches a class-1 PoE implementation with a maximum power consumption of 3.8W will
be aimed for. This requires careful optimisation of the design. A potential option is to use a low power
FPGA (e.g. XILINX Spartan-6) and to distribute the processor functionality onto two embedded
MicroBlaze144 soft processors, one running the fragment management and the other the message
processing. The performance of the two cores will be in the order of 300 MIPS which is roughly one
third of the performance of the current PPC440GP processor. Some additional advantage can be
expected from the improved integration of the MicroBlaze cores with the FPGA fabric. For example,
the access to the FIFO and DPR structures, which is currently a bottleneck, could be done with
customised processor instructions which are available via the MicroBlaze co-processor interface. Due
to the similarity of the MicroBlaze and PowerPC architectures a large fraction of the existing software
sources can be re-used.
143 http://standards.ieee.org/getieee802/download/802.3af-2003.pdf
144 http://www.xilinx.com/publications/prod_mktg/MicroBlaze_Sell_Sheet.pdf
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
109
Chapter 7 - Discussion and Conclusions
7.2.3
Upgrade Phase-2
For the phase-2 of the LHC upgrade significant modifications of the entire TDAQ architecture might
be necessary in order to deal with longer L1 latencies and the increase in event sizes by a factor of 10.
Currently the expectation is that the basic architecture with an L2-trigger operating on RoIs will still
be valid. As a result, a new readout system will be needed, probably integrated with the detector
readout (iROS). The development of the iROS will build on the experiences gained during phase-1.
Two likely options are the move from S-Link to 10GbE as the output interface of the iROS and a
shared memory architecture of the L1 and the ROS fragments.
However, if the RoI principle cannot be pursued for phase-2 it is likely that an architecture similar to
CMS will be used, pushing all L1 accepted data via a fast network (probably 10GbE) into a HLT farm.
7.3 Summary
The ATLAS readout architecture and the ROBIN component have been presented in this thesis. The
design and implementation of the ROBIN component was a success. At the component level, all
performance goals are already meet or at least can be met by tuning of the architecture. System tests
have demonstrated proper operation and good performance, even though the crucial test – operation on
the beam with particle collisions – could not be performed yet due to the delays on the LHC machine.
The expected life time of the present dataflow system is in the order of 10 years, up to the LHC
upgrade phase-2. From the present design of the ROBIN a number of topics have been identified
which need to be modified to achieve higher performance or to support future features of the system.
Prominent examples are the PCIe version of the ROBIN which is already in the development stage,
the design of a ROB-on-ROD architecture for the pending phase-1 upgrade of the pixel sub-detector
and the potential integration of ROD and ROS in the course of upgrade phase-2. It is expected that
many features of the present ROBIN will be migrated via technology upgrades to the new designs.
110
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
8
Appendix
8.1 Pre-ROBIN hardware
8.1.1
MPRACE-1
MPRACE-1 is a PCI based FPGA co-processor developed by the author of this work. It consists
basically of a large XILINX Virtex-2 FPGA, two banks of fast SRAM, one DRAM memory slot, a
commercial PCI bridge and expansion connectors (Figure 62). MPRACE-1 was extensively used in
various research and educational projects, for example as frame-grabber with on-board image
processing [HEZEL1], for prototyping of GE and processor modules for the ROBIN (see chapter
5.3.1.4 ), prototyping of a fast L2 trigger algorithm for the TRT sub-detector of ATLAS [ATLTRT] and
for acceleration of astrophysical simulations [NBODY1].
MPRACE-1 was also used to prototype the PCI communication mechanism used in the bus-based
ROS. Due to the very encouraging results, the implementation was re-used with little modifications on
the final ROBIN. In addition, the software part of the buffer management was implemented directly on
Figure 62: MPRACE-1
the FPGA in order to evaluate a processor-less ROBIN. While the raw functionality could be verified,
complex monitoring and network oriented message passing mechanisms cannot be implemented with
reasonable effort without processor.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
111
Chapter 8 - Appendix
8.1.2
µEnable
The early experiences at the university of Mannheim with FPGA technology led to the development of
the FPGA co-processor µEnable, a PCI card with FPGA, local memory and expansion connectors. The
card (Figure 63) was used in various research projects [MENABLE] and stimulated, together with the
FPGA development tool CHDL [CHDL1][CHDL2], the foundation of the company
SiliconSoftware145, where the author is co-founder.
Figure 63: µEnable
The µEnable card uses a low-density FPGA of the XILINX XC4000 series and a 32 bit PCI interface
and targets rather low-cost applications as compared to MPRACE-1, which provides more features at
higher cost. The card was used as one of the first ROB prototypes, with a ring-buffer capable to store
up to 1.000 fragments. The on-board connectors could be used with standard S-Link or PMC 146
mezzanines.
145 http://siliconsoftware.de/
146 PMC is a standard for PCI mezzanine cards according to IEEE P1386.1
112
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
8.2 ROBIN configuration parameter set
Parameter
Description
sernum
Board serial number, resides in the one-time-programmable area Yes
of the on-baord FLASH memory
BaseIPAddress
Network address, common to all channels.
SubDetectorId
ID of the sub-detector which is connected to the ROBIN. Value is No
inserted into the ROB fragment. Common to all channels.
Pagesize
Buffer memory granularity (page size), defaults to 2kB
Numpages
Number of memory pages, normally auto-calculated from buffer Yes
size and page size.
Hashbits
Number of hash-bits, defaults to 16
Yes
NetworkEnabled
Controls processing of network interface
No
Keepfrags
Deleted fragments are not actually removed from database. Used No
together with certain emulation modes.
Rolemu
Enables data emulation mode
No
TestSize
Size of emulated fragments
No
DebugEnabled
Enables debugging output via “printf” on serial terminal
Yes
DumpRolEnabled
Enable debug output of content of incoming fragments
Yes
Interactive
Enables interactive debugging via serial terminal
Yes
Divclearsize
Patch for incorrect request format in dataflow software
Yes
Keepctlwords
Enables capturing of S-Link framing information
Yes
Dpmcache
Enable caching of message memory area
Yes
RolDataGen
Emulation mode with hardware data generator
No
Macflowctl
Activate flow-control handling on network port
Yes
RolEnabled
Activate fragment processing
No
EbistEnabled
Enable extended build-in self test (BIST)
Yes
Emupattern
Data pattern in emulation mode
Yes
Continbist
Run BIST continuously
Yes
Ignorebist
Continue application even after hard BIST errors
Yes
Rolextloop
Enable link loopback test via external fibre. Defaults to chip Yes
internal loopback.
Mactxdrop
Enable dropping of network response packets if output queue full. Yes
Defaults to no dropping
Mgmtcache
Enable caching of bookkeeping memory
Yes
Dmafifocheck
Enable checking of free space in output queue. Defaults to ON
Yes
Upfdma
Use memory-to-memory DMA to read items from UPF
Yes
Hdrdma
Use memory-to-memory DMA to write DMA headers
Yes
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Expert
No
No
113
Chapter 8 - Appendix
Parameter
Description
Expert
Prescalefrag
Relative inverse priority of fragment processing in main task loop Yes
Prescalemsg
Relative inverse priority of message processing in main task loop Yes
Prescalefpf
Relative inverse priority of FPF processing in main task loop
Yes
UDPBasePort
UDP port number for network responses. Obsolete
No
Max1618check
Enables checking of temperature threshold
Yes
ChannelId
ID of ROBIN channel. Value is inserted into the ROB fragment
No
DcNodeId
ROBIN node id for switch-based ROS mode
No
Secsiemu
Emulate OTP sector for factory testing
Yes
GcLost
Minimum number of lost delete messages to enable garbage No
collection
GcPages
Maximum number of free buffer pages to enable garbage No
collection
MaxRxPages
Threshold for input fragment truncation
No
TempAlarm
Temperature threshold value
Yes
BofNoWait
Patch for incorrect fragment format
Yes
IrqEnable
Enable interrupts to host on error conditions
Yes
DiscardMode
Accept but do not store fragments in stop-less recovery mode
Yes
NetDeleteEnable
Enable processing of network delete messages
No
Table 7: ROBIN configuration parameters
8.3 ROBIN statistic
Statistics item
Description
ERRORS
Hardware errors
Detection of an internal hardware error
Software errors
Detection of a software error condition (coding
error)
Software warning
Detection of a condition which should not occur
(e.g. unexpected code location)
Buffer full occurrences
Transitions into buffer-full state
ROL error
Errors signalled from HW page management
ROL DOWN occurencies
Transitions into link down status
ROL XOFF occurencies
Transitions into link up status
Buffer manager receive errors
Inconsistent event ids on subsequent pages of the
same event
Buffer manager request errors
Inconsistent event database
Temperature warning
Temperature above threshold
114
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
Statistics item
Description
PCI DMA reset
PCI DMA reset due to excessive transaction delay
PCI DMA abort
PCI DMA abort after multiple resets
Interrupts
Interrupts sent to host
FRAGMENTS
Frags received
Fragments received from link
Frags available
Fragments sent to PCI/Network
Frags not available
Fragments requested but not in database
Frags pending
Fragments unavailable but due to arrive
Frags added
Fragments added to database
Frags deleted
Fragments removed from database
Frags truncated
Fragments truncated due to oversized
Frags corrupted
Fragments with soft format or data errors
Frags rejected
Fragments with unrecoverable format errors
Frags replaced
Fragments overwritten due to duplicate event id
Frags out of sync
Mismatch of event id on subsequent fragments
Frags missing on delete
Fragments not in database during delete request
Frags TT sync error
Fragment with “sync” trigger type does not match
on event id and trigger type mask
PAGES
Pages received
Pages received from link
Pages added
Pages added to database
Pages deleted
Pages deleted from database
Pages provided
Pages sent to PCI/Network
Pages supressed
Pages discarded due to fragment truncation
Pages invalid
Pages with format error
Pages dma'ed
Pages received by fragment DMA
MESSAGES
Messages received
Raw messages from PCI/Network
Messages accepted
Valid messages, after decoding
Messages rejected
Invalid format or request code
Messages lost
Lost messages, detected via message sequence
number
Messages invalid
Invalid network format
Messages data request
Requests for data
Messages data response
Responses to data requests
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
115
Chapter 8 - Appendix
Statistics item
Description
Messages clear request
Delete request messages
Messages broadcast
Ethernet broadcasts
Messages PCI TX queue full
PCI response submitted while response queue
occupied. Introduces delay.
Messages NET TX queue full
Network response submitted while response
queue occupied. Introduces delay.
Messages NET TX dropped
Network response dropped due to excessive delay
on output queue
Messages NET RX frames OK
GE MAC statistics: received error-free ethernet
frames
Messages NET RX frames FCS error
GE MAC statistics: received ethernet frames with
CRC error
Messages NET RX frames length error
GE MAC statistics: received ethernet frames with
incorrect length error
Messages NET TX frames OK
GE MAC statistics: error-free ethernet frames
transmitted
Messages NET TX frames underrun
GE MAC statistics: outbound ethernet frames
dropped due to underrun
Messages PCI Tested Empty
Check for new message did not yield request from
PCI
Messages PCI Tested OK
Check for new message provided new request
from PCI
Messages NET Protocol ARP
Network message for ARP
Messages NET Protocol IP
Network message using IP protocol
Messages NET Protocol RS
Network message using raw Ethernet protocol
Messages NET Protocol UDP
Network message using UDP/IP protocol
Messages NET Protocol unknown
Network message using unknown protocol
Messages NET received PAUSE frames
Incoming flow-control message on network
Table 8: ROBIN statistic record
8.4 ATLAS detector parameters
Sub-detector
Channels
ROLs
Event size [kB]
Pixel
1.4 * 108
132
60
Silicon strip (SCT)
6.2 * 106
Inner detector
Transition radiation (TRT)
92
110
5
3.7 * 10
192
307
1.8 * 105
762
576
Calorimeter
LAr calorimeter
116
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
Sub-detector
Channels
ROLs
Event size [kB]
1.0 * 104
64
48
Monitored drift tube (MDT)
3.7 * 105
204
154
Cathode strip chamber (CSC)
6.7 * 104
16
256
3.5 * 105
32
12
5
4.4 * 10
24
6
L1 Calo
NA
48
28
Other
NA
2
0.3
Tile calorimeter
Muon system
L1 Trigger
Resistive plate chamber (RPC)
Thin gap chamber (TGC)
Table 9: ATLAS detector parameters
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
117
Chapter 8 - Appendix
118
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
8.5 Glossary
BERT
CERN
COTS
CP
CPU
CRC
DAQ
DPM
DPR
DRAM
DSP
EB
ECR
EF
FE
FIFO
FPGA
GE
GMAC
GUI
HDL
HLT
iROS
L1
L1ID
L2
L2SUP
L3
LDC
LHC
LSC
LSI
LVDS
MAC
MSI
MIPS
MTU
PHY
QCD
RAM
RoI
RoIB
RoIC
ROS
RTL
SFI
SM
Bit error rate test
European Organisation for Nuclear Research, Geneva, Switzerland
Component of the shelf. In this context used for computers and peripherals designed for
the mass market
Charge/Parity
Central processing unit
Cyclic redundancy check
Data acquisition
Same as → DPR
Dual-ported → RAM
Dynamic → RAM
Digital signal processing
Event building
Event counter reset
Event filter
Fast Ethernet (100Mbit/s)
First-in first-out
Field-programmable gate array
Gigabit-Ethernet
Gigabit-Ethernet media access controller
Graphical user interface
Hardware description language
High-level triggers
Integrated read out system
First level trigger
L1 event identifier, synonymous to event number
Second level trigger
Second level trigger supervisor
Third level trigger
Link destination card (S-Link receiver)
The Large Hadron Collider, a proton-proton particle accelerator built underground at
CERN with a circumference of 27km
Link source card (S-Link transmitter)
Large scale integration
Low voltage differential signalling
Media access controller
Medium scale integration
Million instructions per second
Maximum Transfer Unit (equivalent to Ethernet packet size)
Physical layer adapter
Quantum Chromo Dynamic, a sector of the → SM
Random access memory
Region of interest
Region of interest builder
Region of interest collection
Readout system
Register transfer level – a precise, low-level description of functionality
Switch to farm interface
Standard model of particle physics
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
119
Chapter 8 - Appendix
SRAM
SUSY
TDAQ
TRT
TTL
URD
VHDL
ZBT
120
Static → RAM
Super symmetry, an extension to the → SM
Trigger and data-acquisition system
Transition radiation tracker
Transistor transistor logic
User requirements document
Very high speed hardware description language
Zero bus turn-around → RAM, a variation of synchronous → SRAM
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
List of figures
Figure 1: Experiment Scenario............................................................................................................... 7
Figure 2: TDAQ levels........................................................................................................................... 8
Figure 3: LHC at CERN*..................................................................................................................... 12
Figure 4: CMS Detector [CMSJINST]................................................................................................. 14
Figure 5: CMS DAQ [CMSJINST]...................................................................................................... 15
Figure 6: CMS FED-Builder Rails [CMSJINST]................................................................................. 16
Figure 7: FRL image and block diagram[CMSFRL]............................................................................ 17
Figure 8: ALICE dataflow [ALICEJINST]...........................................................................................17
Figure 9: ALICE D-RORC [DRORC]..................................................................................................18
Figure 10: LHCb dataflow[LHCBADD].............................................................................................. 18
Figure 11: LHCb TELL1 [LHCBTELL]...............................................................................................19
Figure 12: ATLAS detector*.................................................................................................................21
Figure 13: ATLAS TDAQ TP-version [ATLASTP]..............................................................................22
Figure 14: Sequential Selection [SEQSEL].......................................................................................... 24
Figure 15: TDR ROBIN Blockdiagram................................................................................................ 25
Figure 16: Schematic layout of ATLAS TDAQ system [ATLASTDR].................................................26
Figure 17: Dataflow network [STANCU].............................................................................................29
Figure 18: L2 processing times [L2PROC].......................................................................................... 30
Figure 19: Racks with ROS-PCs*.........................................................................................................33
Figure 20: ROS-PC...............................................................................................................................33
Figure 21: PCI load board.....................................................................................................................34
Figure 22: ROBIN TestSuite.................................................................................................................36
Figure 23: ROS-ROBIN interaction..................................................................................................... 37
Figure 24: Virtex-2 logic element......................................................................................................... 42
Figure 25: Virtex-2 FF896 package...................................................................................................... 43
Figure 26: Simulation waveform.......................................................................................................... 46
Figure 27: ChipScope waveform view..................................................................................................47
Figure 28: ROBIN basic hardware elements.........................................................................................57
Figure 29: ROBIN S-Link interface..................................................................................................... 61
Figure 30: Burst-size dependant memory bandwidth............................................................................62
Figure 31: PowerPC performance comparison..................................................................................... 64
Figure 32: PPC440GP mezzanine.........................................................................................................65
Figure 33: DDR memory subsystem.....................................................................................................65
Figure 34: DDR memory simulation.................................................................................................... 66
Figure 35: PPC Xbus device-paced burst write timing......................................................................... 67
Figure 36: PLX local bus timing...........................................................................................................68
Figure 37: JTAG routing.......................................................................................................................70
Figure 38: Final ROBIN....................................................................................................................... 71
Figure 39: ROBIN VHDL modules...................................................................................................... 73
Figure 40: FPGA buffer manager......................................................................................................... 75
Figure 41: FPGA network interface......................................................................................................77
Figure 42: Software task loop...............................................................................................................81
Figure 43: Software buffer manager..................................................................................................... 81
Figure 44: Buffer manager database..................................................................................................... 82
Figure 45: Fragment handling...............................................................................................................83
Figure 46: PCI fragment request and response..................................................................................... 84
Figure 47: Software execution times from profiling............................................................................. 85
Figure 48: Rack wizard.........................................................................................................................89
Figure 49: MTF Entry...........................................................................................................................90
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
121
Chapter 8 - Appendix
Figure 50: Actual PCI stand-alone performance................................................................................... 95
Figure 51: Calculated PCI performance................................................................................................96
Figure 52: ROBIN network performance............................................................................................. 97
Figure 53: Effect of hash-key size........................................................................................................ 98
Figure 54: Garbage collection performance..........................................................................................99
Figure 55: ROS test setup.....................................................................................................................99
Figure 56: Performance of the bus-based ROS...................................................................................100
Figure 57: ROS request rate, emulated data........................................................................................101
Figure 58: Spurious mode oscillation................................................................................................. 102
Figure 59: Temperature sensors.......................................................................................................... 103
Figure 60: PLX unwetted pads........................................................................................................... 104
Figure 61: Bathtub curve.................................................................................................................... 104
Figure 62: MPRACE-1....................................................................................................................... 111
Figure 63: µEnable............................................................................................................................. 112
Figures marked with *: Copyright CERN
List of tables
Table 1: Pilot Project Technologies.......................................................................................................23
Table 2: Data-flow components (TDR).................................................................................................27
Table 3: XILINX FPGA families..........................................................................................................44
Table 4: ROBIN FPGA memory utilisation.......................................................................................... 59
Table 5: FPGA connectivity..................................................................................................................60
Table 6: FPGA resource utilisation....................................................................................................... 78
Table 7: ROBIN configuration parameters......................................................................................... 114
Table 8: ROBIN statistic record..........................................................................................................116
Table 9: ATLAS detector parameters.................................................................................................. 117
122
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
Bibliography
[SMS1] CERN, Introduction to the Standard Model, CERN, 2008,
http://public.web.cern.ch/Public/en/Science/StandardModel-en.html
[SMS2] Herrero, M.J., The Standard Model, Presentation at 10th NATO ASI on Techniques and
Concepts of High-Energy Physics, St. Croix, USA,Jun 1998, arXiv:hep-ph/9812242v1
[DISSFLICK] Flick, T., Studies on the Optical Readout for the ATLAS Pixel Detector , Bergische
Universität Wuppertal, Jul 2006, urn:nbn:de:hbz:468-20060600
[SUSY] Peskin, M., Supersymmetry in Elementary Particle Physics, SLAC,SLAC-PUB-13079, Jan
2008, http://arxiv.org/abs/0801.1928v1
[CPV] Peskin, M., The Matter with antimatter, SLAC,Nature 419,24 - 27, Sep 2002,
http://dx.doi.org/doi:10.1038/419024a
[QCD] Hands, S., The Phase Diagram of QCD, Contemp.Phys.42:209-225,2001. May 2001,
http://arxiv.org/abs/physics/0105022v1
[LHCJINST] Evans, L. et al., LHC Machine, Journal of Instrumentation,JINST 3 S08001, Aug 2008,
http://dx.doi.org/10.1088/1748-0221/3/08/S08001
[ATLJINST] The ATLAS Collaboration, The ATLAS Experiment at the CERN Large Hadron Collider,
Journal of Instrumentation,JINST 3 S08003, Aug 2008, http://dx.doi.org/10.1088/17480221/3/08/S08003
[ALICEJINST] The ALICE Collaboration, The ALICE experiment at the CERN LHC, Journal of
Instrumentation,JINST 3 S08002, Aug 2008, http://dx.doi.org/10.1088/1748-0221/3/08/S08002
[CMSJINST] CMS Collaboration, The CMS experiment at the CERN LHC, Journal of
Instrumentation,JINST 3 S08004, Aug 2008, http://dx.doi.org/10.1088/1748-0221/3/08/S08004
[LHCBJINST] LHCb Collaboration, The LHCb Detector at the LHC, Journal of
Instrumentation,JINST 3 S08005, Aug 2008, http://dx.doi.org/10.1088/1748-0221/3/08/S08005
[DJFROB] D. Francis et al., The Read-Out Buffer in DAQ/EF Prototype -1, CERN,ATL-DAQ-2000053, Aug 2000, http://doc.cern.ch//archive/electronic/cern/others/atlnot/Note/daq/daq-2000053.pdf.
[LHCBDAQ] Alessio, F. et al., LHCb Online event processing and filtering, Journal of Physics,Conf.
Ser. 119 022003, 2008, http://dx.doi.org/10.1088/1742-6596/119/2/022003
[LHCBTB] The LHCb Collaboration, LHCb Technical Proposal, CERN, Feb 1998, http://lhcbtp.web.cern.ch/lhcb%2Dtp/postscript/tp.ps
[CMSTRIG] Afaq, A. et al., The CMS High Level Trigger System, IEEE transactions on nuclear
science,Vol 55 Issue 1, Feb 2008, http://dx.doi.org/10.1109/TNS.2007.910980
[SLINK64] Racz, A. et al, The S-LINK 64 bit extension specification: S-LINK64, CERN, Aug 2003,
http://cmsdoc.cern.ch/cms/TRIDAS/horizontal/docs/slink64.pdf
[CMSSFB] Bauer, G. et al., The Terabit/s Super-Fragment Builder and Trigger Throttling System for
the Compact Muon Solenoid Experiment at CERN, CERN,CERN-CMS-CR-2007-020, May 2007,
http://cdsweb.cern.ch/record/1046342/files/CR2007_020.pdf
[CMSGBE] Bauer, G. et al., CMS DAQ Event Builder Based on Gigabit Ethernet, IEEE Transactions
on Nuclear Science,Vol 55, Issue 1, Feb 2008, http://dx.doi.org/10.1109/TNS.2007.914036
[CMSFRL] Arcidiacono, R., Flexible custom designs for CMS DAQ, Proceedings of the 10th Topical
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
123
Chapter 8 - Appendix
Seminar on Innovative Particle and Radiation Detectors,Vol 172,174 - 177, Oct 2007,
http://dx.doi.org/10.1016/j.nuclphysbps.2007.08.106
[DRORC] Carena, F. et al., The ALICE Data-Acquisition Read-out Receiver card, CERN,Proc. 10th
Workshop on Electronics for LHC and Future Experiments,273ff, Boston, USA,Sep 2004,
http://doc.cern.ch//archive/cernrep/2004/2004-010/p273.pdf
[ALICETDR] ALICE collaboration, Trigger, Data Acquisition, High Level Trigger, Control System
Technical Design Report, CERN, 2004, https://edms.cern.ch/file/456354/2/DAQ_Chapters7-10.pdf
[LHCBADD] Tatsuja, N. et al., Addendum to the LHCb Online System Technical Design Report,
CERN,CERN-LHCC-2005-039, 2005, http://cdsweb.cern.ch/record/903611/files/lhcc-2005039.pdf?version=2
[LHCBTELL] Bay, A. et al., The LHCb DAQ interface board TELL1, Nuclear Instruments and
Methods in Physics Research Section A, Vol 560/2,494ff, May 2006,
http://dx.doi.org/10.1016/j.nima.2005.12.212
[SLINK] Boyle, O., The S-LINK Interface Specification, CERN, Mar 1997,
http://hsi.web.cern.ch/HSI/s-link/spec/spec/
[ATLASTP] The ATLAS collaboration, ATLAS High-Level Triggers, DAQ and DCS: Technical
Proposal, CERN, Mar 2000, http://cdsweb.cern.ch/record/434668/files/cer-2184259.pdf
[ATLDEMPROG] Blair, R. et al., OPTIONS FOR THE ATLAS LEVEL-2 TRIGGER, CERN,OPEN99-149, Feb 1997, http://cdsweb.cern.ch/record/398762/files/open-99-149.ps.gz
[SEQSEL] Bystricky, J., A sequential processing strategy for the ATLAS event selection, IEEE
Transactions on Nuclear Science,Vol 44 Issue 3,342 - 347, Jun 1997,
http://dx.doi.org/10.1109/23.603668
[PAPMOD] Dobson, M. et al., Paper Models of the ATLAS Second Level Trigger, CERN,ATL-DAQ98-113, Jun 1998, http://cdsweb.cern.ch/record/683664/files/daq-98-113.pdf
[PILPRO] Blair, R. et al., The ATLAS Level-2 Trigger Pilot Project, IEEE transactions on nuclear
science,Vol 49 Issue 3,851 - 857, Jun 2002, http://dx.doi.org/10.1109/TNS.2002.1039577
[ROBCPLX] Cranfield, R., Vemeulen, J., Options for the ROB Complex, CERN,ATL-DAQ-2000027, Apr 2000, http://cdsweb.cern.ch/record/684047/files/daq-2000-027.pdf
[AROBC] Boeck, R. et al., The active ROB complex, CERN,ATL-DAQ-2000-022, Mar 2000,
http://cdsweb.cern.ch/record/683960/files/daq-2000-022.pdf
[ATLASTDR] Jenni, P. et al., ATLAS high-level trigger, data-acquisition and controls : Technical
Design Report, CERN, Jul 2003, http://cdsweb.cern.ch/record/616089/files/cer-002375189.pdf
[STANCU] Stancu, S. N., Networks for the ATLAS LHC Detector : Requirements, Design and
Validation, CERN,CERN-THESIS-2005-054, Jul 2005,
http://cdsweb.cern.ch/record/913894/files/thesis-2005-054.pdf
[GETB] Matei Ciobotaru, Stefan Stancu, Micheal LeVine, and Brian Mart, GETB, a GigabitEthernet
Application Platform: its Use in the ATLAS TDAQ Network, IEEE transaction on nuclear
science,Vol 53 Issue 3,817 - 825, Jun 2006, http://dx.doi.org/10.1109/TNS.2006.873303
[ETHERT] Barnes, F.R.M. et al., Testing ethernet networks for the ATLAS data collection system,
IEEE transactions on nuclear science,Vol 49 Issue 2,516 - 520, Apr 2002,
http://dx.doi.org/10.1109/TNS.2002.1003791
[BASEDF] Beck, H.-P. et al., The Base-Line DataFlow System of the ATLAS Trigger and DAQ, IEEE
transactions on nuclear science,Vol 51 Issue 3,470 - 475, Jun 2004,
http://dx.doi.org/10.1109/TNS.2004.828707
124
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
[LARSC] Burckhart-Chromek, D. et al., Testing on a Large Scale : running the ATLAS Data
Acquisition and High Level Trigger Software on 700 PC Nodes, Proceedings of 15th International
Conference on Computing In High Energy and Nuclear Physics,60 - 65, Mumbai, India,Mar2006,
http://cdsweb.cern.ch/record/941077
[ETHER] S. Stancu, R.W. Dobinson, M. Ciobotaru, K. Korcyl, and E. Knezo, The use of Ethernet in
the Dataflow of the ATLAS Trigger and DAQ, CERN,Proceedings of Computing in High Energy
and Nuclear Physics, La Jolla, CA, USA,Mar 2003, http://arxiv.org/pdf/cs.ni/0305064
[DFNET] Stancu, S. et al., ATLAS TDAQ DataFlow Network Architecture Analysis and Upgrade
Proposal, Proceedings of NPSS Real Time Conference,Vol 53 Issue 3,826 - 833, Jun 2005,
http://dx.doi.org/10.1109/TNS.2006.873302
[DFROS] Vermeulen, J.C. et al., ATLAS DataFlow : the Read-Out Subsystem, Results from Trigger
and Data-Acquisition System Testbed Studies and from Modeling, Proceedings of IEEE NPSS Real
Time Conference,Vol 10 Issue 10, Jun 2005, http://dx.doi.org/10.1109/RTC.2005.1547446
[L2PROC] Abolins, M. et al., Integration of the Trigger and Data Acquisition Systems in ATLAS,
Journal of Physics,Conf. Ser. 119 (2008) 022001, 2008, http://dx.doi.org/10.1088/17426596/119/2/022001
[MADROS] Müller, M., ROS Architecture and Requirements, CERN, May 2004,
http://agenda.cern.ch/fullAgenda.php?ida=a041681
[ATLDCS] Bariusso Poy, A. et al., The detector control system of the ATLAS experiment, Journal of
Instrumentation,JINST 3 P05006, May 2008, http://dx.doi.org/10.1088/1748-0221/3/05/P05006
[ROSURD] Cranfield, R., LeVine, M., McLaren, R., ROS User Requirements Document, CERN,ATLDQ-ES-0007, May 2004, https://edms.cern.ch/file/356336/1.0.2/ros_urd_v102.pdf
[ROBROD] Beck, H.P., Hauser, R., LeVine, M., Impact of a ROB on ROD Design on Atlas
DataCollection, CERN,ATL-DQ-ER-0002 , Nov 2001, https://edms.cern.ch/file/391562/0.5/DC027.pdf
[ROBRODC] Beck, H.P. et al., Commissioning with a ROB-on-ROD Based Readout, CERN,ATLDQ-ON-0001 , Mar 2003, https://edms.cern.ch/file/374790/0.1/RoRcommissioning_v01.pdf
[XLNXDS31] XILINX, XILINX Virtex-2 Datasheet, XILINX,Vendor Datasheet, May 2005,
http://www.xilinx.com/support/documentation/data_sheets/ds031.pdf
[OSI] ISO, Open systems interconnect reference model, ISO,ISO/IEC 7498-1, Jun 1996,
http://standards.iso.org/ittf/PubliclyAvailableStandards/s020269_ISO_IEC_7498-1_1994%28E
%29.zip
[ETH802.3] IEEE, IEEE 802.3, IEEE, Dec 2005, http://standards.ieee.org/getieee802/802.3.html
[ROBSUM] ATLAS ROS Group, ROBIN Summary, CERN, Feb 2002,
http://atlasinfo.cern.ch/Atlas/GROUPS/DAQTRIG/ROS/documents/ROBINsummary.pdf
[HLDDPROT] Green, B., Kugel, A., Kieft, G., Prototype-RobIn HLDD, CERN, Sep 2002,
https://edms.cern.ch/file/356324/2.4/hldd.pdf
[DLDDPROT] Green, B., Kugel, A., Kieft, G., Protoype-RobIn DLDD, CERN, Sep 2002,
https://edms.cern.ch/file/356328/2.3/dldd.pdf
[SWIDPROT] Green, B., Kugel, A., Kieft, G., Protoype-RobIn Software Interface, CERN, Sep 2002,
https://edms.cern.ch/file/356332/2.2/swid.pdf
[FDRPROT] Farthouat, P., Report of the Final Design Review ROBIN Prototype, CERN, Oct 2002,
https://edms.cern.ch/file/359014/1/robin-pdr.pdf
[ROBRDD] Green, B. et al., ROBIN Design Document, CERN, May 2004,
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
125
Chapter 8 - Appendix
https://edms.cern.ch/file/473396/1.0/rdd.pdf
[ROBFDR] Farthouat, P., Report of the Final Design Review ROBIN, CERN, Aug 2004,
https://edms.cern.ch/file/478897/2/Robin-fdr-may2004.pdf
[ROBPRR] Farthouat, P., PRR of the ROBIN, CERN, Mar 2005,
https://edms.cern.ch/file/571432/1/Robin-prr-mar2005.pdf
[PRRDD] Green, B. et al., ROBIN Production Readiness Review: Design Document, CERN, Jan
2005, https://edms.cern.ch/file/555100/1/prr_design_050221.pdf
[PRRPM] Green, B. et al, , CERN, Jan 2005,
https://edms.cern.ch/file/555103/1/prr_performance_050221.pdf
[PRRTP] Green, B. et al., ROBIN Production Readiness Review: Test Procedures, CERN, Jan 2005,
https://edms.cern.ch/file/555102/1/prr_testing_050221.pdf
[PRRPS] Green, B. et al, ROBIN Production Readiness Review: Production Schedule, CERN, Jan
2005, https://edms.cern.ch/file/555104/1/prr_procurement_050222.pdf
[ROBSPARE] Kieft, G. et al., ROBin Spares Policy, CERN, Mar 2006,
https://edms.cern.ch/file/714623/1.3/Robin_Spares_v1-3.pdf
[SWROB] Iwanski, W. et al., The software ROBIN, CERN, Feb 2002,
http://indico.cern.ch/getFile.py/access?
resId=1&materialId=0&contribId=s1t14&sessionId=s1&subContId=0&confId=a0281
[MMROB] Müller, M., Evaluation of an FPGA and PCI Bus based Readout Buffer for the Atlas
Experiment, Mannheim University, May 2005, http://madoc.bib.unimannheim.de/madoc/volltexte/2005/1070/
[GBELSC] Blair, R. et al., A Gigabit Ethernet Link Source Card, Argonne national laboratory, 2002,
http://lhc-electronics-workshop.web.cern.ch/LHC-electronics-workshop/2002/DAQ/B33.pdf
[ROBROD] Beck, H.P., Hauser, R., LeVine, M., Impact of a ROB on ROD Design on Atlas
DataCollection, CERN,ATL-DQ-ER-0002 Nov 2001, https://edms.cern.ch/file/391562/0.5/DC027.pdf
[POE] IEEE, IEEE Std. 802.3af - IEEE Standard for Information technology - Telecommunications
and information exchange between systems - Local and metropolitan area networks - Specific
requirements, IEEE,Standard for Information technology 802.3af, 2003,
http://ieeexplore.ieee.org/servlet/opac?punumber=8612
[VME] IEEE Computer Society, IEEE Standard for A Versatile Backplane Bus: VMEbus,
IEEE,ANSI/IEEE Std 1014-1987, 1987, http://dx.doi.org/10.1109/IEEESTD.1987.101857
[UKROB] Boorman, G. et al., The UK ROB-in, a prototype ATLAS readout buffer input module,
CERN,ATL-DAQ-2000-013, Mar 2000, http://cdsweb.cern.ch/record/684041/files/daq-2000013.pdf
[ROBMAN] Kugel, A. et al., ATLAS ROBIN User Manual, CERN,ATL-DQ-ON-0018, Apr 2006,
https://edms.cern.ch/file/719553/1/robinUserManual.pdf
[DC022] Beck, H.-P. et al., The Message Format for the ATLAS TDAQ DataCollection, CERN,ATLDQ-ES-0035 , Oct 2008, https://edms.cern.ch/document/391557/2.5
[RAWFMT] dos Anjos, A. et al., The raw event format in the ATLAS Trigger & DAQ, CERN,ATL-DES-0019 , Feb 2009, https://edms.cern.ch/document/445840/4.0c
[ATLCOM] Morettini, P. et al., ATLAS Detector Commissioning, CERN,CERN-ATL-SLIDE-2008178, Nov 2008, http://cdsweb.cern.ch/record/1140714/files/ATL-SLIDE-2008-178.pdf
126
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
[ROBJINST] Cranfield, R. et al., The ATLAS ROBIN, Journal of Instrumentation,JINST 3 T01002,
Jan 2008, http://dx.doi.org/10.1088/1748-0221/3/01/T01002
[TIPP09] Crone, G. et al., The ATLAS ReadOut System - performance with first data and perspective
for the future, Acc. for publication in proceedings of The 1st international conference on
Technology and Instrumentation in Particle Physics, Tsukuba, Japan,Mar 12-172009,
[ATLRDY] Vandelli, W. et al., Readiness of the ATLAS Trigger and Data Acquisition system for
thefirst LHC beams, 11th Topical Seminar on Innovative Particle and Radiation Detectors,ATLCOM-DAQ-2009-005 Siena, Italy,Oct 01-04 2008,
http://cdsweb.cern.ch/record/1155455/files/ATL-COM-DAQ-2009-005.pdf
[ROSRT09] Della Volpe, D. et al., The ATLAS DataFlow System: Present Implementation,
Performance and Future Evolution, Subm. for presentation at RealTime Conference 2009, Bejing,
China,May 2009,
[HEZEL1] Hezel, S., FPGA-basiertes Template-Matching mit Distanztransformierten Bildern,
Mannheim University, Jul 2004, http://madoc.bib.uni-mannheim.de/madoc/volltexte/2004/338/
[ATLTRT] Hinkelbein, C. et al, Using of FPGA Coprocessor for Improving the Execution Speed of the
Pattern Recognition Algorithm for ATLAS – High Energy Physics Experiment, Lecture Notes in
Computer Science, Vol 3203/2004,791-800, Aug 2004,
http://www.springerlink.com/content/kj9hg110eadf2vjd/
[NBODY1] Spurzem, R et al., From Newton to Einstein – N -Body Dynamics in Galactic Nuclei and
SPH using new special hardware and Astrogrid-D, Journal of Physics,Conf. Ser. 78 012071 2007,
http://dx.doi.org/10.1088/1742-6596/78/1/012071
[MENABLE] Brosch, O. et al, MICROENABLE - A RECONFIGURABLE FPGA COPROCESSOR,
CERN,Proc. 4th Worksh. on Electronics for LHC Experiments,402ff Rome, Italy,1998,
[CHDL1] Kornmesser, K. et al., The FPGA Development System CHDL, IEEE,Proc. of the 9th IEEE
FCCM conference,271 - 272, Napa, CA,Apr 2001, ISBN: 0-7695-2667-5
[CHDL2] Kornmesser, K., The FPGA Development System CHDL, Mannheim University, Dec 2004,
urn:nbn:de:bsz:180-madoc-8575
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
127
Chapter 8 - Appendix
128
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
Chapter 8 - Appendix
Acknowledgements
I wish to express my gratitude to all the people who helped me during this work, first of all to my
supervisor Prof. Männer, who supported me with patience and advice in many respects and stayed
confident in my skills to complete this thesis.
I thank my spiritual mentor Ursa Paul, who encouraged me and helped me to stay focused on this task.
My wife and my children helped me to keep my spirit up, in particular in the last period when putting
all things together. Thanks to you!
This work would not have been possible without the collaboration with and contributions from former
and present colleagues from my institute and from the other ROS groups - my thank goes to David
Francis for getting this project on the way and to Benedetto Gorini for the good guidance of the overall
ROS activity; to Jos Vermeulen for looking deeply into models and measurement results; to Markus
Joos and Louis Tremblet for their help during all CERN on-site activities; to Barry Green and Gerard
Kieft, Matthias Müller and Erich Krause for their friendly and efficient cooperation during the
hardware and firmware developments; to Andrzej Misiejuk, Stefan Stancu and Hans-Peter Beck for
their help setting the networking stuff up and to Nicolai Schroer, Per Werner and Kostas Kordas for
running so many tests with it.
I also express my thanks to all the others who I didn't mention by name but who helped with ideas,
hints and discussions or just with caring about the progress of this work.
The ATLAS ROBIN – A High-Performance Data-Acquisition Module
129
Download PDF