User manual | Resource-Efficient Real-Time Scheduling Using Credit-Controlled Static-Priority Arbitration

Resource-Efficient Real-Time Scheduling
Using Credit-Controlled Static-Priority Arbitration
Firew Siyoum, Benny Akesson, Sander Stuijk, Kees Goossens, Henk Corporaal
Eindhoven University of Technology, Department of Electrical Engineering,
email:{f.m.siyoum, k.b.akesson, s.stuijk, k.g.w.goossens, h.corporaal}@tue.nl
Abstract—A present-day System-on-Chip (SoC) runs a wide
range of applications with diverse real-time requirements. Resources, such as processors, interconnects and memories, are
shared between these applications to reduce cost. Resource
sharing causes temporal interference, which must be bounded
by a suitable resource arbiter. System-level analysis techniques
use the service guarantee of the arbiter to ensure that realtime requirements of these applications are satisfied. A service
guarantee that underestimates the minimum service provided by
an arbiter results in more allocation of resources than needed
to satisfy latency and throughput requirements. For instance,
a linear service guarantee cannot accurately capture bursty
service provision by many priority-based schedulers, such as
Credit-Controlled Static Priority (CCSP) and Priority-Budget
Scheduling (PBS). As a result, the timing analysis of these arbiters
becomes too pessimistic. This leads to unnecessary cost penalties
since some SoC resources, such as SDRAM bandwidth, are scarce
and expensive.
This paper addresses this problem for the CCSP arbiter.
The two main contributions are: (1) a piecewise linear service
guarantee that accurately captures bursty service provisioning,
and (2) an equivalent dataflow model of the new service guarantee, which is an essential component to integrate the arbiter
with dataflow-based system-level design techniques that analyze
the worst-case latency and throughput of real-time applications.
The new service guarantee enables efficient resource utilization
under CCSP arbitration. Experimental results of an H.263 video
decoder application show that memory bandwidth savings from
26% up to 67% can be achieved by using the new service
guarantee as compared to the existing linear service guarantee.
(PBS) [3] allow low latency to bursty requestors, which are
users whose request rate fluctuate significantly over time.
A linear service guarantee based on the latency-rate (LR)
server model [4] cannot capture the bursty service provision
by these arbiters. Therefore, it provides a pessimistic WCRT
that leads to exaggerated latency and diminished throughput.
As a consequence, a designer has to over-allocate resources
to satisfy latency and throughput requirements. This may
overshadow the benefits of these arbiters for SoC resource
sharing. CCSP, in particular, has three key properties that
make it suitable for scheduling access to shared MPSoC
resources [2], [5]: (1) it bounds temporal interference and
guarantees each requestor a minimum allocated service in
a time interval, (2) it has an efficient allocation mechanism
that enables high resource utilization, and (3) it has a small
area hardware implementation that is also fast enough to keep
up with the speed of SoC resources and make a scheduling
decision every clock cycle. As shown in Figure 1, the current
service guarantee of CCSP, referred to as a latency-rate service
guarantee, is a linear lower bound on the provided service.
However, like many priority-based arbiters, a requestor under
CCSP can have bursty provided service intervals, as illustrated
in the figure. As a result, the latency-rate service guarantee
gives a pessimistic WCRT for CCSP arbitration. System-level
timing analysis based on this service guarantee, consequently,
results in over-allocation of SoC resources.
I. I NTRODUCTION
Convergence of application domains in consumer electronics requires a system to run a wide range of applications with
diverse real-time requirements. To satisfy the computational
requirements of these applications at low power consumption,
sophisticated Multi-Processor Systems-on-Chip (MPSoC) with
heterogeneous processing elements are used [1]. To reduce
cost, system resources, such as processors, interconnects and
memories, are shared between multiple users, which we refer
to as requestors. However, resource sharing causes temporal
interference between requestors, which must be bounded by
a resource arbiter to guarantee a certain amount of service to
each requestor. The minimum service an arbiter guarantees a
requestor in an interval is referred to as its service guarantee.
This guarantee bounds the worst-case response times (WCRT)
of requestors, which is used by system-level analysis techniques to determine if real-time requirements, such as latency
and throughput, of applications are satisfied.
A service guarantee that underestimates the minimum service provided by an arbiter results in more allocation of resources than needed, which we refer to as over-allocation. For
instance, priority-based schedulers, such as Credit-Controlled
Static Priority (CCSP) [2] and Priority-Budget Scheduling
Accumulated
service units
Index Terms—real-time; SoC; CCSP; service guarantee; over
allocation; arbitration; dataflow; latency-rate server
requested
service
provided
service
improvement
in WCRT
bursty provided
service interval
new
bi−rate
guarantee
existing
latency−rate
guarantee
Time [service cycle]
Fig. 1.
New piecewise linear vs. existing linear service guarantee.
In this work, we show how to address the over-allocation
problem through a piecewise linear service guarantee. Piecewise linear bounds have been previously used for trafficshaping [6]–[8], as well as computing response time of
arbiters [9]. However, these works focus on regulating unbounded requested services to make service guarantee analysis
feasible. On the other hand, our model is a piecewise linear
lower bound on the worst-case provided service. The model
is an improved service guarantee that accurately characterizes
the provided service. Most importantly, this work shows how
a piecewise linear service guarantee can be represented with
an equivalent dataflow model. This brings the service guarantee from individual access requests to the application level.
Ultimately, this allows us to compute the worst-case latency
and throughput of applications that run on MPSoCs using
dataflow-based system-level design techniques [10], [11].
This paper has two main contributions: (1) a piecewise
linear service guarantee for CCSP, which we refer to as birate service guarantee, as shown in Figure 1, and (2) an
equivalent dataflow model of the bi-rate service guarantee. As
a result of the new bi-rate model, a given SoC resource under
CCSP arbitration can support more requestors, or a given set of
requestors can be accommodated with less resource capacity.
Experiments on traced traffic of an H.263 video decoder show
that we can save from 27% up to 67% of memory bandwidth,
compared to the existing latency-rate service guarantee.
The remaining part of this paper is organized as follows. In
Section II, a review of related work is presented. Section III
recaps the CCSP arbiter and presents the formal model used
in this paper. The new bi-rate service guarantee and the associated dataflow model, which are the two major contributions
of this paper, are presented in Sections IV and V, respectively.
Section VI covers the experimental setup and obtained results.
Finally, the paper concludes in Section VII, summarizing the
highlights of this work.
II. R ELATED WORK
Simple and starvation-free schedulers are natural candidates
to arbitrate resource accesses in real-time MPSoCs. This is
because they provide upper bounds on temporal interference
and have small and fast hardware implementations. Several
of such arbiters have been proposed, including TDMA [12],
and extensions of Round-Robin arbitration, such as Weighted
Round-Robin [13], and Deficit Round-Robin [14]. These arbiters offer requestors a fixed allocated budget (rate), which is
replenished within a fixed frame size (period). These framebased arbiters guarantee each requestor a minimum service
proportional to its allocated rate. However, they fail to efficiently support diverse latency requirements without overallocating resources. For instance, they cannot provide low
latency to a requestor with low bandwidth requirement without
allocating more slots in a frame. This problem is referred to
as coupling between latency and rate [15]. Over-allocation in
these types of arbiters may also result from a similar coupling
between allocation granularity and latency. For instance, a
single slot in a TDMA table with four entries corresponds to
25% of the bandwidth. Over-allocation occurs if the requestor
requires any less. The over-allocation can be reduced by
increasing the number of entries, although, this increases the
WCRT of all requestors sharing the resource [2], [15]. Priority
Budget Scheduling (PBS) [3] addresses the coupling between
latency and rate using priorities. However, it only improves
the latency of a single bursty requestor with high priority
and still couples allocation granularity and latency. On the
other hand, efforts to decouple the allocation granularity from
latency have been made in [16]–[18], although the resulting
arbiters couple latency and rate, and are unable to efficiently
distinguish diverse latency requirements.
CCSP circumvents the shortcomings of existing resource
arbiters, since (1) it efficiently supports diverse latency requirements, and (2) offers a small hardware implementation that
runs sufficiently fast and keep up with resources like SDRAM
memory controllers [19]. It has furthermore been shown in [2]
that CCSP belongs to the class of LR servers, and hence offers
a minimum guaranteed service to its requestors in an interval.
This service guarantee is used at system-level to determine
if latency and throughput requirements of applications are
satisfied. However, the latency-rate service guarantee is a
linear lower bound that cannot capture the bursty service
provided by many priority-based schedulers. As a result, it
yields a pessimistic WCRT for requestors under CCSP. This
ultimately leads to unnecessarily excess resource allocations
to attain the latency and throughput requirements.
This paper presents a piecewise linear service guarantee for
CCSP that captures the bursty provided service by modeling
two distinct modes in the arbiter, namely when a requestor
has enough budget for a bursty service, and when it does not.
Our analysis is based on service curves, commonly used in
analysis of timing bounds and buffer requirements in the field
of communication networks [20]–[23]. A common similarity
between these works is their use of traffic-shaping to enforce
a certain request arrival curve. Piecewise linear bounds have
also been employed for traffic-shaping/policing in [6]–[8], as
well as for computing response time of real-time arbiters by
bounding resource access requests [9]. However, these traffic
shapers are applied to the requested service, whereas our
model is a piecewise lower-bound on the provided service.
Dataflow graphs allow analysis of the worst-case latency
and throughput of real-time systems [24], [25]. State-of-theart dataflow-based system-level analysis techniques [10], [11]
require various aspects of the system, such as computation,
buffers and arbitration, to be modeled with dataflow components. Modeling resource sharing with dataflow components is
not trivial, as equivalence in temporal behavior with the service
guarantees of the arbiters should be proved using rigorous
algebraic steps [26], [27]. In [28], a general dataflow model
for arbiters in the class of LR servers is presented that also
covers CCSP. However, the basis of this dataflow model is
the current linear LR service guarantee, and hence suffers
from the associated over-allocation problem for priority-based
arbiters. In [29], a three-actor dataflow model is presented
for the PBS arbiter. However, this model only targets shared
memory access and the latency improvement applies only to
the single high-priority requestor. In this paper, we present a
new dataflow model for CCSP based on our new piecewise
linear service guarantee that applies to any MPSoC resource
as well as to requestors with any priority level.
III. BACKGROUND
The CCSP arbiter was originally proposed in [2]. A small
and fast hardware implementation of the arbiter was presented
in [5]. In this section, we iterate the basic operation of the
CCSP arbiter and key definitions that are essential for a complete presentation of this work. In Section III-A, the formal
model used in this paper is discussed. In Section III-B and
III-C, the basic operation of the CCSP arbiter and the current
service guarantee are explained using this formal model.
A. Formal Model
The analysis of the CCSP arbiter uses service curves [20] to
model the interaction between the requestors and the resource.
We use an abstract resource view where a service unit is
the access granularity of the resource. Time is discrete and
a time unit, referred to as a service cycle, is defined as the
time required to serve such a service unit. We use closed
discrete time intervals and hence [τ, t] includes all cycles
in the sequence hτ, τ + 1, ..., t − 1, ti. A service curve is a
cumulative function of service units. For any service curve
ξ, ξ(t) denotes its value at service cycle t. We furthermore use ξ(τ, t) = ξ(t + 1) − ξ(τ ) to denote the difference
in values between the endpoints of the closed interval [τ, t].
The set of requestors is denoted as R. The k th request
from requestor r ∈ R is ωrk and its size in service units is
s(ωrk ) ∈ N+ . Arriving requests from each requestor to the
resource are placed in separate buffers in front of the resource.
The arriving requests from a requestor are captured by the
requested service curve, w, as defined in Definition 1.
Definition 1 (Requested service curve): The requested service curve of a requestor r ∈ R is denoted wr (t) : N → N,
where wr (0) = 0 and
wr (t + 1) =
wr (t) + s(ωrk )
wr (t)
ωrk arrived at t + 1
no request arrived at t + 1
The scheduled requestor at time t is denoted
γ(t) : N → R ∪ {∅} and receives service of one service
unit. The service provided by the resource to a requestor
is captured by the provided service curve, w′ , as given in
Definition 2.
Definition 2 (Provided service curve): The provided service curve of a requestor r ∈ R is denoted wr′ (t) : N → N,
where wr′ (0) = 0 and
′
wr (t) + 1 γ(t) = r
′
wr (t + 1) =
wr′ (t)
γ(t) 6= r
At any time t the difference between the requested service
curve and the provided service curve gives us the number of
service units waiting in the buffer for service. This is called the
backlog of the requestor, defined in Definition 3. The concepts
of requested service curve, provided service curve and backlog
are illustrated in Figure 2.
Definition 3 (Backlog): The backlog of a requestor r ∈ R
at any time t is denoted qr (t) : N → N, and is defined as
qr (t) = wr (t) − wr′ (t).
Accumulated
service units
w
q(τ )
w′
s(ω k )
τ
the allocated burstiness (σr′ ) and the allocated rate (ρ′r ), as
given in Definition 4. These two parameters, in combination,
determine the fraction of resource capacity that the requestor
is allocated.
Definition 4 (Allocated service): The service allocation of
a requestor r ∈ R is the pair (σr′ , ρ′r )P∈ R+ × R+ . For
′
a valid allocation, it should hold that
∀r∈R ρr ≤ 1 and
′
∀r ∈ R : σr ≥ 1.
CCSP uses a continuous budget replenishment policy, which
is based on the concept of active periods, which is defined in
Definition 5. Intuitively, the active period of a requestor is the
maximum interval of time where (1) it is backlogged and/or
(2) it is live, as defined in Definition 6. A requestor is said
to be live at time t if the cumulative of the requested service
units is not less than the total service the requestor would get
if it were continuously getting service at its allocated rate. A
requestor in its active period interval is said to be active and
Rta denotes the set of active requestors at time t.
Definition 5 (Active period): An active period of a requestor r ∈ R is defined as the maximum time interval [τ1 , τ2 ],
such that ∀t ∈ [τ1 , τ2 ] : qr (t) > 0 ∨ wr (τ1 − 1, t − 1) ≥
ρ′r · (t − τ1 + 1).
Definition 6 (Live requestor): A requestor r ∈ R is defined as live at a time t during an active period [τ1 , τ2 ] if
wr (τ1 − 1, t − 1) ≥ ρ′r · (t − τ1 + 1).
Potential refers to the amount of budget a requestor has
and it is defined in Definition 7. The allocated burstiness,
σ ′ , determines the initial potential of a requestor at the start
of an active period, while the allocated rate, ρ′ , corresponds
to the speed with which the budget is replenished during
every service cycle in an active period. Together these two
parameters determine the upper bound on the provided service of a requestor, ŵ′ , as illustrated in Figure 3. A high
allocated burstiness entitles a requestor to more service before
exhausting its potential, forcing it to surrender the resource to
lower priority requestors. For every service unit a requestor
is provided, the potential is decremented by one. Figure 3
illustrates the relationships between allocated service (ρ′ , σ ′ ),
potential (π(t)) and active period ([τ1 , τ2 ] and [τ3 , τ4 ]).
Definition 7 (Potential): The potential of a requestor r ∈ R
is denoted πr (t) : N → R, where πr (0) = σr′ and

πr (t) + ρ′r − 1
πr (t + 1) = πr (t) + ρ′r
 ′
σr
Time [service cycle]
Requested and provided service curves and related concepts.
B. CCSP Arbiter
A CCSP arbiter that arbitrates a set of requestors R sharing
a common system resource consists of a rate regulator and a
scheduler [2]. The scheduler uses a static-priority scheduling
scheme. Each requestor is assigned a unique priority level
and the set of requestors that have higher priority than a
requestor r ∈ R is denoted Rr+ . In addition, each requestor
has an allocated service that consists of two parameters:
live line
Accumulated
service units
Fig. 2.
r ∈ Rta ∧ γ(t) = r
r ∈ Rta ∧ γ(t) 6= r
r∈
/ Rta ∧ γ(t) 6= r
upper bound on
provided service
w
w′
π(t)
ρ′
σ′
τ1
t
τ2
τ3
τ4
Time [service cycle]
Fig. 3.
Allocated service, potential and active period.
A requestor is said to be eligible for scheduling: (1) if it is
backlogged i.e. qr (t) > 0 and (2) if it has enough potential for
at least one service unit i.e. πr (t) ≥ 1 − ρ′r (since the potential
of an active requestor increments by ρ′r every service cycle).
The set of eligible requestors at time t is denoted Rte . The
scheduler of CCSP is non-work-conserving and schedules the
highest priority eligible requestor every service cycle.
C. Latency-rate Service Guarantee
Previous work [2] on the service guarantee of the CCSP
arbiter shows that a requestor is guaranteed service according
to its allocated rate, ρ′ , after a maximum latency, Θ. This
linear service guarantee defines a lower bound, w̌′ , on the
provided service curve during an active period (cf. Figure 4).
This lower bound, which we refer to as the latency-rate
service guarantee, shows that CCSP belongs to the class of
LR servers [4] and it is used to derive the WCRT of requests.
Based on this latency-rate service guarantee, an active
requestor r ∈ R is guaranteed a minimum service during
an active period [τ1 , τ2 ] according to the following relation:
∀t ∈ [τ1 , τ2 ] : w̌r′ (τ1 , t) = max(0, ρ′r · (t − τ1 + 1 − Θr )),
where
X
σs′
Θr =
∀s∈Rr+
1−
X
ρ′s
(1)
∀s∈Rr+
IV. A RBITER A NALYSIS FOR B I - RATE S ERVICE
G UARANTEE
In this section, we present a piecewise linear service guarantee for the CCSP arbiter. The service guarantee of an
arbiter is the minimal service it can provide to a requestor,
irrespective of the situation of other requestors that are sharing
the resource. This minimal service is normally computed by
considering the worst-case scenario that leads to the WCRT.
For a requestor under the CCSP arbiter, this worst-case scenario happens when it experiences the maximum interference
from higher priority requestors. This is stated in Lemma 1 and
proven in [2].
Lemma 1 (Maximum interference): The maximum interference experienced by a requestor r ∈ R during an interval
[τ1 , τ2 ] occurs when all higher priority requestors start an
active period at τ1 and remain active ∀t ∈ [τ1 , τ2 ], and equals
X
σs′ + ρ′s · (τ2 − τ1 + 1)
(2)
îr (τ1 , τ2 ) =
∀s∈Rr+
As discussed in Section III-C, the existing service guarantee
of CCSP guarantees a requestor r ∈ R its allocated rate, ρ′r ,
after a maximum latency, Θr . However, in CCSP a requestor
can be temporarily served at a rate higher than its allocated rate
after its maximum latency. Intuitively, this can be explained as
follows. At the end of the maximum latency, a requestor can
have an accumulated potential from two sources, according to
Definition 7: (1) from its allocated burstiness i.e. if σr′ > 1 and
(2) from potential accumulated while being blocked by higher
priority requestors, i.e. during the maximum latency, Θr . As
a result, at the end of the maximum latency, a requestor has
a potential that is equal to the sum of these two.
When a requestor is at the end of its maximum latency,
it implies that higher priority requestors have utilized their
accumulated potential. Thus, they have to accumulate potential at their respective allocated rate during multiple service
cycles, before they are eligible to access the resource again.
Consequently, the requestor can use the resource whenever
it is not used by higher priority requestors. This means, the
requestor can get service at a higher rate ρ∗r ≥ ρ′r , where
X
ρ′s .
(3)
ρ∗r = 1 −
∀s∈Rr+
The requestor receives service at this higher rate as long
as its potential does not go below 1 − ρ′r , which is the
minimum potential a requestor needs to have to be eligible
for scheduling. The service cycle at which the potential drops
below 1 − ρ′r is referred to as the boundary cycle. After the
boundary cycle the requestor has to wait multiple service
cycles and accumulate potential at its allocated rate to be
eligible. Therefore, it receives service at its allocated rate, ρ′r .
If the potential does not drop below 1 − ρ′r throughout the
active period, the boundary cycle is considered to be the last
service cycle of the active period. In this case, the requestor
receives all its service units in the active period with the higher
service rate, ρ∗r .
Definition 8 (Boundary Cycle): The boundary cycle of a
requestor r ∈ R is defined as the maximum service cycle τrb
during an active period [τ1 , τ2 ] of the worst-case scenario, such
that
∀t ∈ [τ1 , τrb ] : πr (t) > 1 − ρ′r .
(4)
The boundary cycle can be determined based on Definition 8, as shown in Lemma 2. Note that in this paper, we
only consider requestors for which ρ∗r > ρ′r , since otherwise
the higher service rate is equal to the allocated rate. The case
where ρ∗r = ρ′r can occur only P
to the lowest priority requestor
in a fully loaded resource i.e. ∀r∈R ρ′r = 1.
Lemma 2: For a requestor r ∈ R with active period
[τ1 , τ2 ] of the worst-case scenario, the boundary cycle
τrb = min(τ2 , ⌊tx ⌋) where
X
σs′
σr′ −1+ρ′r +
(5)
∀s∈Rr+
t x = τ1 +
∗
′
ρ −ρ
r
r
Proof: The potential of a requestor r at time t ∈ [τ1 , τ2 ]
is given in Equation (6). The term σr′ + ρ′r · (t − τ1 ) is the
potential the requestor gets and wr′ (τ1 , t − 1) is the potential it
spends for the service units it receives in the interval [τ1 , t−1],
according to Definition 7.
πr (t) = σr′ + ρ′r · (t − τ1 ) − wr′ (τ1 , t − 1).
(6)
The service provided to requestor r during the worst-case
scenario is given by Equation (7), where îr is the maximum
interference, given by Lemma 1. It is computed by subtracting
the maximum interference from higher priority requestors
from the total service units available in the interval.
wr′ (τ1 , t − 1) = t − τ1 − îr (τ1 , t − 1)
(7)
Since the higher service rate of CCSP is not captured in
the latency-rate service guarantee, the timing analysis results
in resource over-allocation to satisfy real-time requirements
of requestors. To solve this problem, we propose a piecewise
linear service guarantee, which we refer to as bi-rate service
guarantee. The bi-rate service guarantee takes the higher service rate into account and improves the WCRT of requests, as
shown in Figure 4. ∆ in the figure illustrates the improvement
by using a bi-rate service guarantee.
and the service guarantee is determined solely by the higher
ˇr′ (τ1 , t) = w̌r′h (τ1 , t). In contrast, for
rate linear equation, i.e. w̌
b
′a
t > τr , w̌r (τ1 , t) is the minimum and the service guarantee
ˇr′ (τ1 , t) = w̌r′a (τ1 , t).
is determined by the allocated rate, i.e. w̌
If τ2 , i.e. the end of the active period, is less than tx , then the
requestor gets service at its higher rate throughout the active
period.
w̌r′h (τ1 , t)
ρ∗r
w
w
ρ′
τb
Fig. 4.
τ2
Improving WCRT by piecewise linear model.
Under the bi-rate service guarantee, a requestor r is guaranteed a minimum service at two different rates, ρ∗r and ρ′r ,
after a maximum latency Θr . These two rates correspond to
the cases when a requestor has enough potential to access
the resource at a high rate, and when it does not. The
maximum latency, Θr , is the same as the one in the latencyrate service guarantee, which is given in Equation (1). The
bi-rate guarantee is defined based on two linear equations:
higher-rate guarantee w̌r′h and allocated-rate guarantee w̌r′a ,
which are given in Equations (8) and (9), respectively.
w̌r′h (τ1 , t) = ρ∗r · (t − τ1 + 1 − Θr )
(8)
w̌r′a (τ1 , t)
(9)
=
ρ′r
where
Γr = −
· (t − τ1 + 1 − Γr )
σr′ + ρ∗r − 1
ρ′r
The new bi-rate service guarantee is, then, the minimum of
the two at any time. Hence, during an active period [τ1 , τ2 ], a
ˇr′ (τ1 , t),
requestor r ∈ R is guaranteed a minimum service, w̌
as given in Equation (10).
ˇr′ (τ1 , t) = max(0, min(w̌r′h (τ1 , t), w̌r′ (τ1 , t)))
w̌
a
Fig. 5.
Higher-rate and allocated-rate guarantee.
Time [service cycle]
τrh = τrb − Θr .
Time [service cycle]
Θ
Θr
To guarantee a provided service at the higher rate, ρ∗ ,
the requestor has to ask for service at this rate. Thus, the
new service guarantee applies only to requestors that have a
minimum service request rate of ρ∗ at least up to time point
τ h (cf. Figure 4), given in Equation (11). We refer to such
requestors as high-rate requestors, defined in Definition 9.
ρ∗
τh
τ1
Γr
′
Accumulated
Service Units
∆
ρ′r
τrb
τ1
live line
w̌′ (latency-rate)
ˇ′ (bi-rate)
w̌
w̌r′a (τ1 , t)
Accumulated
Service Units
Putting Equation (7) into Equation (6), gives us the complete
relation for πr (t). Solving for tx from the relation πr (tx ) =
1 − ρ′r results in Equation (5). Then the boundary cycle τrb
is the largest integer less than tx , since time is discrete, and
does not exceed τ2 , which is the end of the active period. This
gives us the relation τrb = min(τ2 , ⌊tx ⌋).
(10)
Γr is computed such that the intersection of the two linear
equations, w̌r′h (τ1 , t) and w̌r′a (τ1 , t), is at tx , given in Equation (5). Hence, solving the two linear equations simultaneously shows that their intersection point is, indeed, at tx .
If τ2 , which is the end of the active period, is greater than
this intersection point tx , then the boundary cycle τrb = ⌊tx ⌋
according to Lemma 2. Therefore, as illustrated in Figure 5,
for t ≤ τrb , w̌r′h (τ1 , t) is the minimum of the two equations
(11)
Definition 9: (High-rate requestor) A requestor r ∈ R
is a high-rate requestor during an active period [τ1 , τ2 ]
if ∀t ∈ [τ1 , τ2 ]:
∗
t ≤ τrh
ρr · (t − τ1 + 1)
wr (τ1 , t) ≥
wr (τ1 , τrh ) + ρ′r · (t − τrh ) Otherwise
Next, we present the mathematical proof of this new service
guarantee, but first two important lemmas are provided to simplify this proof. Equation (6) shows an interesting relationship
between the provided service and potential. By rearranging, we
arrive at Lemma 3.
Lemma 3: For a requestor r ∈ R during an active period
[τ1 , τ2 ], it holds that ∀t ∈ [τ1 , τ2 ]
πr (t) ≤ σr′ − ρ′r ⇐⇒ wr′ (τ1 , t − 1) ≥ ρ′r · (t − τ1 + 1) (12)
From Lemma 3, it has also been proved in [2] that the
eligibility of a requestor can be determined by looking at its
potential, as given in Lemma 4.
Lemma 4: For a requestor r ∈ R during an active period
[τ1 , τ2 ], it holds that ∀t ∈ [τ1 , τ2 ]
πr (t) > σr′ − ρ′r =⇒ r ∈ Rte .
(13)
Theorem 1 defines the bi-rate service guarantee, making use
of the concepts and results presented so far.
Theorem 1: (Bi-rate service guarantee) During any active
period [τ1 , τ2 ] and ∀t ∈ [τ1 , τ2 ], a high-rate requestor r ∈ R
is guaranteed a minimum service according to Equation (10).
Proof: The strategy of the proof is showing that the
service guarantee holds for four different types of time intervals [τi , τj ] that cover all possible scenarios during the active
period.
Case 1: ∀t ∈ [τi , τj ] ∧ τj ≤ τrb : qr (t) > 0. This case covers
time intervals before the boundary cycle where the requestor
is backlogged. For this case, we want to show that a minimum
service is guaranteed at the higher rate of the requestor; i.e.
ˇr′ (τ1 , t) = w̌r′h (τ1 , t). Since we are before the boundary
w̌
cycle, we know that ∀t ∈ [τi , τj ], πr (t) > 1 − ρ′r , according
to Definition 8. This means the requestor is eligible, since the
definition of this case states that it is also backlogged. The
resource provides (τj − τi + 1) service units in the interval.
Taking into account the maximum interference possible from
higher priority requestors (cf. Lemma 1), the minimum service
available to r is given as:
ˇr′ (τi , τj ) = τj − τi + 1 − îr (τi , τj ).
w̌
(14)
Expanding and rearranging the right side of Equation (14)
ˇr′ (τi , τj ) = ρ∗r · (τj − τi + 1 − Θr ), proving a minimum
gives w̌
service at the higher rate for all timing intervals of this type.
Case 2: ∀t ∈ [τi , τj ] ∧ τj ≤ τrb : qr (t) = 0. This case covers
time intervals before the boundary cycle where the requestor
is not backlogged. For time intervals of this type, we want
to prove that a minimum service can be guaranteed at the
higher rate. By Definition 3, the backlog of a requestor is
the difference between the requested service and the provided
service. If there is no backlog, this means the requested and
provided services are equal i.e. wr (t) = wr′ (t). Since r
is a high-rate requestor, it holds that the requested service
wr (τ1 , t) ≥ ρ∗r · (t − τ1 + 1) for the interval [τ1 , τrh ]. Therefore,
for this interval, wr′ (τ1 , t) ≥ ρ∗r ·(t−τ1 +1) ≥ w̌r′h (τ1 , t). Since
wr (t) is a non-decreasing function, it also holds for the interval
[τrh , τrb ] that wr′ (τ1 , t) ≥ w̌r′h (τ1 , t). Therefore, w̌r′h (τ1 , t) is the
guaranteed service.
Case 3: ∀t ∈ [τi , τj ] ∧ τi > τrb : πr (t) > σr′ − ρ′r . This
case considers intervals after the boundary cycle where the
requestor has more potential than its allocated burstiness. After
the boundary cycle, the requestor is no more in the higher rate
service interval of the active period. Therefore, in this interval
we want to prove that a minimum service can be guaranteed
ˇr′ (τ1 , t) = w̌r′a (τ1 , t). Since πr (t) >
at its allocated rate, i.e. w̌
′
′
σr − ρr , the requestor is eligible according to Lemma 4.
Similar arguments as in Case 1 of this proof regarding the
available service units and the maximum interference give the
guaranteed service as ρ∗r · (τj − τi + 1 − Θr ), which equals
w̌r′h (τi , τj ). For t > τrb , w̌r′a (τi , τj ) is less than w̌r′h (τi , τj ).
Therefore, w̌r′a (τi , τj ) can be taken as the guaranteed service.
Case 4: ∀t ∈ [τi , τj ] ∧ τi > τrb : πr (t) ≤ σr′ −
ρ′r . This case considers time intervals after the boundary cycle where the potential of the requestor is less
than its allocated burstiness. The strategy of the proof is
first to show that wr′ (τ1 , t − 1) ≥ w̌r′a (τ1 , t). Then, based
on Definition 2 of provided service curve, it follows that
wr′ (τ1 , t) ≥ wr′ (τ1 , t − 1) ≥ w̌r′a (τ1 , t).
1) The provided service wr′ (τ1 , t − 1) can be expanded as
shown in Equation (15), since for any service curve ξ,
ξ(τ, t) = ξ(t + 1) − ξ(τ ).
wr′ (τ1 , t − 1) = wr′ (τ1 , τrb − 1)+
wr′ (τrb , τrb ) + wr′ (τrb + 1, t − 1).
(15)
We proceed by bounding the three terms in Equation (15) individually.
2) At τrb , requestor r is scheduled, because its potential drops, which happens only when a requestor gets
service, according to Definition 7. This implies that
wr′ (τrb + 1) = wr′ (τrb ) + 1. Using this, wr′ (τrb , τrb ) can be
formulated as follows.
wr′ (τrb , τrb ) = wr′ (τrb + 1) − wr′ (τrb ) = 1
3) From Case 1 and 2 of this proof, we already proved that
wr′ (τ1 , τrb − 1) ≥ ρ∗r · (τrb − τ1 − Θr )
In addition, computing the intersection of the two linear
equations ρ∗r · (τrb − τ1 − Θr ) and ρ′r · (τrb − τ1 ) + σr′ − 1
shows that for ∀t > τrb it holds that
wr′ (τ1 , τrb − 1)
≥ ρ∗r · (τrb − τ1 − Θr )
≥ ρ′r · (τrb − τ1 ) + σr′ − 1.
4) Since πr (t) ≤ σr′ − ρ′r we have from Lemma 3 that
wr′ (τrb + 1, t − 1) ≥ ρ′r · (t − τrb )
5) Now putting, the results from steps 2, 3 and 4 into
Equation (15), we get:
wr′ (τ1 , t − 1)
≥ 1 + ρ′r · (τrb − τ1 ) +
σr′ − 1 + ρ′r · (t − τrb )
= 1 + ρ′r · (t − τ1 ) + σr′ − 1
≥ ρ′r + ρ∗r + ρ′r · (t − τ1 ) +
σr′ − 1 (Since 1 > ρ′r + ρ∗r )
= ρ′r · (t − τ1 + 1) + ρ∗r + σr′ − 1
= ρ′r · (t − τ1 + 1 − Γr )
ˇr′ (τ1 , t) =
Hence, we can guarantee a minimal service w̌
ρ′r · (t − τ1 + 1 − Γr ) = w̌r′a (τ1 , t).
In summary, the bursty service a requestor can receive
depends mainly on its allocated burstiness and the allocated
burstiness of its higher priority requestors. Increasing its allocated burstiness increases its potential at the start of the active
period. Increasing the allocated burstiness of higher priority
requestors enables the requestor to accumulate more potential
at its allocated rate while being blocked by them. High
accumulated potential means a requestor can receive service at
a high rate, as it does not have to wait to accumulate potential
to be eligible. The bi-rate service guarantee integrates this fact
through the boundary cycle parameter, τ b , and guarantees each
requestor high-rate service until its boundary cycle for every
active period. This results in an improvement in the WCRT
of requests as compared to latency-rate service guarantee that
does not consider this bursty provided service. The reduction
in the WCRTs of individual requests lead to improvements
in latency and throughput of applications. Since designers
are concerned in satisfying these real-time requirements, a
system-level representation of the service gurantee is needed.
Dataflow-based system-level design technique is one efficient
approach to achieve this goal. In the next section, we present
a dataflow model of the bi-rate service guarantee.
A. Dataflow-based System-level Analysis
A DFG [32] is a directed graph that consists of actors that
are connected through channels. A channel represents a FIFO
buffer through which actors communicate by sending tokens.
A token represents an abstract data unit. The connection point
between an actor and a channel is referred to as a port. An
example DFG consisting of two actors X, Y and two channels
cxx , cxy is shown in Figure 6(a). Channels that have the same
source and destination actor are referred to as self-edges (e.g.
channel cxx ). A number next to a black dot over a channel
represents the number of initial tokens that are available in the
channel at the beginning of the execution of the graph.
A variety of DFGs exist today with different levels of analyzability, expressiveness and implementation efficiency [33].
Homogeneous Synchronous Dataflow graph (HSDFG) is one
of them. In HSDFG, when there exists at least one token on
every input port of an actor, the actor fires. At the end of
the firing, the actor produces one token on each of its output
ports. The execution duration of a single firing of an actor Y
is denoted χ(Y ) ∈ R+ . Depending on the system aspect a
given actor models, the implication of the execution duration
may vary. For instance, for an actor that models a software
process, the execution duration may represent its worst-case
execution time (WCET) for a given processor type.
Figure 6(b) shows an example binding of actors X and Y
onto an example MPSoC platform. The platform is a multi-tile
architecture, where each tile comprises a processing unit (P),
a local memory (M) and a network interface (NI).
By introducing dataflow components that model various
sorts of architectural aspects, such as computation, storage,
arbitration and communication, into the DFG, an architectureaware DFG can be created. A simple example of an
Tile 1
P
1
cxx
M
Tile 2
X
P
NI
Y
M
n1
NI
X
cxy
N
Communication Network
cxy
Memory
Controller
Y
(a) DFG
Fig. 6.
X
1
n2
Shared
Memory
Y
1
(b) Binding of actors
(c) Architectureaware DFG
Use of DFGs for system-level design and analysis of MPSoCs
architecture-aware HSDFG is shown in Figure 6(c). In this
graph the initial tokens n1 and n2 model the allocated buffer
sizes for the channel cxy on Tile 1 and Tile 2, respectively.
The inter-tile communication latency between actors X and Y
is modeled by the execution time of actor N , χ(N ).
Timing analysis and scheduling of this system can be
carried out on the architecture-aware DFG. For example, the
worst-case throughput of the application is the inverse of
the maximum cycle mean (MCM) of the graph, which is the
largest average execution duration among all cycles in the
graph [24]. The approach presented in this section can also
be applied to complex graphs that use more expressive DFGs.
Detailed discussions on dataflow-based system-level design
and analysis techniques can be found in [10], [11] and [33].
B. Dataflow Model for CCSP arbitration
We present a HSDFG model, which is shown in Figure 7,
for service provision using CCSP arbitration for a requestor
r ∈ R. The dataflow model consists of three actors: latency actor (Lr ), higher-rate actor (Hr ) and allocated-rate
actor (Ar ). The number of initial tokens hr and the execution
durations of the three actors are determined by the requestor’s
allocated service and the allocated services of its higherpriority requestors, Rr+ . The number of initial tokens hr is
given in Equation (16), where sr , provided in Equation (17),
refers to the number of service units that can be served at the
higher rate during every active period.
k
j
′
r
(16)
hr = sr − (sr −2)·ρ
ρ∗
r
sr =
j
Θr −Γr
1
− ρ1∗
ρ′
r
r
k
(17)
1
Accumulated
service units
V. DATAFLOW M ODEL
A dataflow graph (DFG) is a directed graph that allows
high-level modeling and analysis of real-time applications.
These graphs have efficient analysis techniques to compute
throughput and buffer requirements of applications that are
running on MPSoCs. They can capture cyclic data dependencies between processes of an application, which exist in many
real-life systems. Because of these reasons, DFGs play an
indispensable role in today’s MPSoC design flows that carry
out application binding and resource scheduling [12], [30].
In addition, dataflow design tools enable fast design-spaceexploration (DSE) of MPSoCs to find arbiter configurations
and other system parameters that satisfy latency and throughput requirements of applications [10], [11], [27].
State-of-the-art dataflow-based system-level analysis techniques require that the various aspects of the system, such
as computation, storage and arbitration, to be modeled with
dataflow components. Dataflow modeling of arbitration is
not trivial, as equivalence in temporal behavior with the
service guarantees of the arbiters should be proved using
rigorous algebraic steps [26]–[29]. In this section, we present
a dataflow model for CCSP arbitration, based on the bi-rate
service guarantee previously derived in Section IV. First in
Section V-A, we briefly introduce how DFGs can be used
for system-level design and analysis of MPSoCs. Section V-B
then presents the dataflow model for the CCSP arbiter, along
with the intuition behind its execution. Due to space reasons,
the complete mathematical proof of the dataflow model is not
presented in this paper, but it is provided in [31].
Ar
hr
Lr
ρ′r
sr
Hr
ρ∗r
Θr
1
(a) CCSP dataflow model
Fig. 7.
Time [service cycle]
(b) Rate of provided service
Dataflow model for CCSP arbitration.
The execution durations of the three actors for a requestor
r ∈ R are given in Equations (18), (19) and (20). The
case where hr = 1 is unique from all other cases (hr > 1),
since this implies that actors Hr and Ar cannot be executed
in parallel. This is because there can exist a maximum of
only one token at a time over their cyclic dependency. The
execution duration of actor Ar , χ(Ar ) for hr = 1 is hence
different from the other cases, as shown in Equation (20).
χ(Lr ) = Θr
(18)
1
ρ∗r
(19)
χ(Hr ) =
χ(Ar ) =
(
1
ρ′r
1
ρ′r
−
1
ρ∗
r
if hr > 1
if hr = 1
(20)
Figure 8 demonstrates the use of the CCSP dataflow model
in the context of system-level design by extending Figure 6(c)
to model CCSP arbitrated shared memory communication
between actors X and Y of Figure 6(a). Actor N of Figure 6(c)
is replaced by two CCSP dataflow models for the write and
read latencies of actors X and Y . Actors Nx and Ny model
the latencies over the communication network.
1
1
Ax
Ay
hy
hx
X
Nx
Lx
Hx
Ny
Ly
1
1
Hy
Y
1
1
n2
n1
token by actor Hr marks the finishing time of the j th service
unit, denoted F(j). F(j) is determined by the arrival time of
tokens from its three input channels, shown in Figure 7(a).
These are: (1) the arrival of the j th token from the output of
actor Lr , E(j) + Θr , (2) the finishing time of the previous
firing, F(j − 1), which models that only one service unit can
be served at a time and (3) the production time of the (j −
hr )th token from actor Ar . This is because there are already
hr initial tokens in this channel, representing the number of
service units that can be served at the higher rate, ρ∗r , before
the potential is exhausted at the boundary cycle. For every
service unit served at the higher rate, one token is consumed
from these initial tokens. As long as there are tokens available
on this channel, service units can be served at the higher rate.
This is because in order to fire, actor Hr does not have to
wait for the production of tokens by actor Ar . However, after
all tokens are consumed, the rate of production of tokens by
actor Hr is determined by the rate of production of tokens
by the actor Ar . According to Equation (20), this implies that
service units are served at the allocated rate of the requestor,
ρ′r . Notice that for hr = 1, the rate of token production by
actor Hr is the sum of the execution durations of both actors
Hr and Ar . This is because the two actors cannot be fired
simultaneously, since there can exist a maximum of only one
token over the cyclic dependency at a time.
Bringing the above three limiting factors together, the
finishing time of service units can be expressed with a maxplus equation [34] that bounds the completion time of service
units, as shown in Equation (21). G(j) denotes the finishing
time of the j th token at the output of the allocated-rate actor.
cxy
Fig. 8.
Architecture-aware DFG with shared memory communication.
The rate of service provision by the CCSP dataflow model,
shown in Figure 7(b), is equivalent to the bi-rate service
guarantee, previously given in Equation (10). In the remaining
part of this section, the intuition behind the rate of service
provision of the dataflow model is explained.
Arriving requests at the input of the CCSP arbiter can be
multiple service units long, i.e. for the k th request s(ω k ) ≥ 1.
Each service unit is represented by a single token in our
dataflow model. The arrival of a request at the input of the
arbiter is analogous to the arrival of multiple tokens at the
input of actor Lr . This means the arrival time of the j th token
at actor Lr , denoted E(j), marks the arrival time of the j th
service unit at the input of the arbiter. Actor Lr produces
a single token for every arriving token after an execution
duration of Θr service cycles, which models the maximum
waiting time every service unit encounters. The production
time of the j th token by the latency actor equals E(j) + Θr .
It is important to notice that actor Lr does not have a selfedge with initial tokens. The number of initial tokens on a
self-edge determines the auto-concurrency of the actor, which
is the number of possible simultaneous firings of the actor.
This implies that multiple tokens can wait in parallel, which
accurately reflects what actually happens at the input of the
arbiter. On the other hand, actors Hr and Ar have self-edges,
each with a single initial token that implies only one service
unit can be served at a time by the resource.
Output tokens produced by actor Hr represent completed
service units. This means that the production time of the j th

max(E(j) + Θr ,


F(j − 1),
F(j) =
G(j − h)) +


0
1
ρ∗
r
j>0
Otherwise
(21)
The mathematical proof of equivalence between Equation (21) and the bi-rate service guarantee, previously given
in Equation (10), is provided in [31].
VI. E XPERIMENT R ESULTS
This section presents experimental results that demonstrate
the implications of our new bi-rate service guarantee for a
real-life application. First, in Section VI-A, the experimental
setup is presented. Then, Section VI-B shows improvements
in guaranteed throughput by applying the bi-rate service guarantee. Finally, Section VI-C demonstrates how the throughput
improvement translates to resource savings when satisfying
real-time requirements.
A. Experimental Setup
Our experimental setup emulates a MPSoC with a SRAM
memory controller that uses the CCSP arbiter to arbitrate
memory accesses from multiple requestors. We limit the number of requestors to five, r1 to r5 , and they are assigned unique
priority levels in decreasing order, giving r1 the highest and
r5 the lowest priorities. Each requestor ri has an associated
allocated rate (ρ′ri ) and allocated burstiness (σr′ i ).
In real-time system design, the resource allocation of requestors should guarantee that each of them satisfies their
timing requirements in the worst-case interference scenario.
Priority
assignment
Memory controller
models
(ri )
1
r1
H.263 decoder
memory access
trace
−clock cycle
−number of
service units
...
L ri
Sr i
w̌r′ i
r2
w
5
7
x 10
6
5
4
3
2
Requested Service Curve
ˇr′ )
Bi−rate Service Guarantee (w̌
3
1
0
0
1
r3
and the application is assigned priority level 3 in the CCSP
arbiter. We return to the experiment with these parameters
later. The modified requested service curves (according to both
service guarantees) are not shown in this figure, since they
ˇr′ ) at this graph
overlap with the two lower bounds (w̌r′ 3 and w̌
3
scale. This graph shows the improvement in latency by using
the bi-rate service guarantee, compared to the latency-rate rate
service guarantee.
Accumulated Service [service unit]
The objective of this experiment is to demonstrate improvements in throughput and resource allocation by applying the
bi-rate service guarantee as compared to the existing latencyrate service guarantee. For this purpose, a MATLAB module
is written that models a simple SRAM memory controller with
a 32-bit interface, both according to the latency-rate dataflow
model of [28] and our new dataflow model, presented in
Section V. Given a requested service curve, wri , the module
computes two lower bounds on the provided service curve,
ˇr′ , based on the latency-rate and the bi-rate dataflow
w̌r′ i and w̌
i
models, respectively. The experiment setup is illustrated in
Figure 9, where the execution duration of actor Sri of the LR
server model is given by χ(Sri ) = 1/ρ′ri .
′
Latency−rate Service Guarantee (w̌r3 )
1
2
3
4
5
6
Time [clock cycle]
7
8
9
7
x 10
Ari
r4
hri
L ri
Hr i
ˇr′
w̌
i
r5
1
Fig. 9.
The experimental setup.
A requested service curve w is generated by tracing memory
transactions of a real-life application. The SimIt-ARM [35]
cycle-accurate simulator is used to trace all external memory
transactions (cache misses) from a StrongArm processor that
runs an H.263 video decoder application (written by Telenor
research, a popular public-domain implementation of H.263).
The simulator is configured with a 16 Kbyte instruction cache
with 32-byte blocks and a 16 Kbyte write-back data cache with
32-byte blocks. In addition, the memory reading and writing
latencies of the simulator are set to zero while generating the
trace. The service unit of the memory controller module is
set to 32-bytes (the cache line) and the service cycle to 8
clock cycles. In the experiment, it is assumed that all memory
read and write transactions stall the processor. The stalling of
the processor is taken into account by modifying the requested
service curve w in such a manner that the arrival time of every
transaction (requested service unit) is delayed by the response
time of the previous transaction. This is important as the trace
is originally generated by setting the write and read latencies
of the simulator to zero.
B. Throughput Improvement
In this experiment, we intend to demonstrate differences in
guaranteed throughput according to the latency-rate and bi-rate
service guarantees. Assuming the traced traffic represents the
worst-case memory access pattern, we compute the throughput
by taking the inverse of the total time required to decode one
complete frame. This gives us the minimum throughput in
frames per second (fps), since the worst-case memory access
pattern yields the maximum total time duration to decode one
frame. Figure 10 shows the service curves while the system
decodes one complete frame where ∀i ρ′ri = 0.15, ∀i σr′ i = 2
Fig. 10. Requested service (generated with zero memory access latency)
and lower bounds for the provided service based on latency-rate and bi-rate
service guarantees.
Table I shows the improvements in throughput as the
assigned priority of the application is varied from 1 to 5.
The allocated rate is fixed at 0.15 for all cases and the
allocated burstiness of all requestors is varied together from 1
to 4. The result shows that requestors with average priority
improve more than both high and low priority requestors.
This is because high-priority requestors have smaller Θ, which
implies that they accumulate less potential while being blocked
and low-priority requestors have smaller ρ∗ relative to highpriority requestors (cf. Equation (3)). Though it is not shown
here, we are also able to observe from this experiment that
the achievable throughput improvements for different allocated
rates have similar trends as the results shown in Table I.
TABLE I
T HROUGHPUT IMPROVEMENTS IN %.
Priority
1
2
3
4
5
∀i σr′ i = 1
27
36
44
41
24
∀i σr′ i = 2
30
48
45
28
15
∀i σr′ i = 3
30
54
36
21
11
∀i σr′ i = 4
31
53
30
17
9
It is important to note that these achieved improvements are
neither at the cost of additional latency penalty of lower priority requestors nor by additional allocated resource capacity.
It is a direct result of the tighter bi-rate service guarantee that
accurately models the service provided by the CCSP arbiter.
C. Efficient Resource Utilization
The improvements in throughput can be traded for resource
savings, since the throughput requirement of a real-time requestor can now be satisfied with less allocated resources.
Assume that our application runs on a 266 MHz processor
and has a throughput requirement of 15 fps. Figure 11 shows
all combinations of service allocations and assigned priorities
that satisfy this requirement, based on both the latency-rate
and bi-rate guarantees.
Latency−rate model
Bi−rate model
0.9
0.8
Resource saving
0.3
Priority 1
Priority 2
Priority 3
0.28
Priority 1
Priority 2
Priority 3
90
Priority 1
Priority 2
Priority 3
80
0.7
0.26
70
0.6
0.5
0.4
resource saving (%)
allocated rate
allocated rate
0.24
0.22
0.2
60
50
40
0.3
30
0.18
0.2
20
0.16
0.1
0
0 1 2 3 4 5 6 7 8 9 1011
allocated burstiness
(a) Latency-rate model
0.14
10
0 1 2 3 4 5 6 7 8 9 1011
allocated burstiness
0 1 2 3 4 5 6 7 8 9 1011
allocated burstiness
(b) Bi-rate model
(c) Resource saving
Fig. 11. Resource allocation for throughput-constrained application: a) shows
latency-rate cannot capture burstiness and b) shows bi-rate can.
It can be seen that by using the bi-rate service guarantee
and without allocating any additional burstiness (keeping
σr′ i = 1), the resource allocation can be reduced by 67% (from
ρ′r = 0.68 to ρ′r = 0.22) at priority level 3, by 43% at priority
level 2 and by 26% at priority level 1. The resource savings
further increase when increasing the allocated burstiness. For
instance, at priority level 3 and σr′ 3 = 10, the resource
allocation can be reduced by 72% (from 0.68 to 0.19). These
improvements are achieved without imposing any additional
latency on lower-priority requestors, since the arbiter configuration is the same in both lower-bound computations.
This experiment shows that the latency-rate service guarantee cannot capture bursty provided service and, hence, fails
to completely decouple latency and rate. In this use case,
the allocated rate must be increased at all priority levels to
satisfy the throughput requirement of 15 fps. This results in
substantial over-allocation of the resource. For instance, when
the application is assigned priority level 3, 46% of the memory
bandwidth is wasted due to the pessimism of the latency-rate
model, which allocates 68% of the bandwidth. In contrast, the
bi-rate service guarantee, presented in this paper, accurately
models the behavior of the CCSP arbiter and accomplishes this
at 22%, and consequently, enables efficient resource utilization
by avoiding over-allocation of the resource.
VII. C ONCLUSIONS
The Credit-Controlled Static-Priority (CCSP) arbiter is suitable for scheduling shared System-on-Chip resources, such as
memories. The existing linear service guarantee of the arbiter
fails to capture that service may be provided in a bursty
manner, resulting in over-allocation of resources to satisfy realtime requirements. In this paper, we address this drawback
through a new piecewise linear (bi-rate) service guarantee. In
addition, for system-level analysis, we present a novel dataflow
model of the arbiter, based on the new service guarantee. As
a result of the new bi-rate guarantee, a given resource under
CCSP arbitration can support more requestors, or a given
set of requestors can be accommodated with less resource
capacity. Experimental results on traced traffic of an H.263
video decoder application show that savings from 26% up to
67% in memory bandwidth can be achieved by using the new
bi-rate guarantee compared to the existing linear guarantee.
R EFERENCES
[1] “International Technology Roadmap for Semiconductors (ITRS),” 2009.
[2] B. Akesson et al., “Real-Time Scheduling Using Credit-Controlled
Static-Priority Arbitration,” in Proc. RTCSA, 2008.
[3] M. Steine et. al, “A priority-based budget scheduler with conservative
dataflow model,” in DSD, 2009.
[4] D. Stiliadis and A. Varma, “Latency-rate servers: a general model for
analysis of traffic scheduling algorithms,” IEEE/ACM Trans., 1998.
[5] B. Akesson et al., “Efficient Service Allocation in Hardware Using
Credit-Controlled Static-Priority Arbitration,” in Proc. RTCSA, 2009.
[6] A. Francini and F. Chiussi, “Minimum-latency dual-leaky-bucket shapers
for packet multiplexers: Theory and implementation,” in IWQoS, 2000.
[7] H. J. Chao and J. S. Hong, “Design of an ATM shaping multiplexer
with guaranteed output burstiness,” Computer Systems Science and
Engineering, 1996.
[8] F. Guillemin et. al, “Extremal traffic and bounds for the mean delay of
multiplexed regulated traffic streams,” in INFOCOM, 2002.
[9] S. Stein et. al, “A polynomial-time algorithm for computing response
time bounds in static priority scheduling employing multi-linear workload bounds,” in ECRTS, 2010.
[10] A. Bonfietti et. al, “An efficient and complete approach for throughputmaximal SDF allocation and scheduling on multi-core platforms,” in
DATE, 2010.
[11] C. Lee et. al, “A systematic design space exploration of MPSoC based
on synchronous data flow specification,” Journal of Signal Processing
Systems, Springer, 2010.
[12] A. Hansson et. al, “CoMPSoC: A template for composable and predictable multi-processor system on chips,” ACM TODAES, vol. 14, no. 1,
2009.
[13] M. Katevenis et. al, “Weighted round-robin cell multiplexing in a
general-purpose ATM switch chip,” IEEE J. Sel. Areas Commun., vol. 9,
no. 8, Oct. 1991.
[14] M. Shreedhar and G. Varghese, “Efficient fair queueing using deficit
round robin,” in Proc. SIGCOMM, 1995.
[15] H. Zhang, “Service disciplines for guaranteed performance service in
packet-switching networks,” Proc. IEEE, vol. 83, no. 10, 1995.
[16] C. Kalmanek et. al, “Rate controlled servers for very high-speed
networks,” Proc. GLOBECOM, 1990.
[17] S. S. Kanhere and H. Sethu, “Fair, efficient and low-latency packet
scheduling using nested deficit round robin,” High Performance Switching and Routing, 2001 IEEE Workshop on, 2001.
[18] D. Saha et al., “Carry-over round robin: a simple cell scheduling
mechanism for ATM networks,” IEEE/ACM Trans. Netw., vol. 6, no. 6,
1998.
[19] B. Akesson et. al, “Predator: a predictable SDRAM memory controller,”
in Proc. CODES+ISSS, 2007.
[20] J.-Y. L. Boudec and P. Thiran, Network calculus: a theory of deterministic queuing systems for the Internet. Springer-Verlag New York, Inc.,
2001.
[21] R. Cruz, “A calculus for network delay. I. Network elements in isolation,” IEEE Trans. Inf. Theory, vol. 37, no. 1, 1991.
[22] G. Kesidis and T. Konstantopoulos, “Worst-case performance of a buffer
with independent shaped arrival processes,” IEEE Communications
Letters, 2000.
[23] E. Wandeler et. al, “Performance analysis of greedy shapers in real-time
systems,” in DATE, 2006.
[24] A. Ghamarian et. al, “Throughput analysis of synchronous data flow
graphs,” in ACSD, 2006.
[25] M. Geilen and S. Stuijk, “Worst-case performance analysis of synchronous dataflow scenarios,” in CODES/ISSS, 2010.
[26] M. Bekooij et. al, “Performance guarantees by simulation of process,”
in SCOPES, 2005.
[27] M. Bekooij et. al, “Dataflow analysis for real-time embedded multiprocessor system design,” in chapter 15. Dynamic and Robust Streaming
Between Connected CE Devices. Kluwer Academic Publishers, 2005.
[28] M. Wiggers et. al, “Modelling run-time arbitration by latency-rate
servers in dataflow graphs,” in SCOPES, 2007.
[29] J. Staschulat and M. Bekooij, “Dataflow models for shared memory
access latency analysis,” in EMSOFT, 2009.
[30] A. Kumar et. al, “Multi-Processor System-Level Synthesis for Multiple
Applications on Platform FPGA,” in FPL, 2007.
[31] F. Siyoum et. al, “Dataflow Model for Credit-Controlled Static-Priority
Arbitration,” http://www.es.ele.tue.nl/esreports/, TU/Eindhoven, Tech.
Rep., 2010.
[32] E. Lee and D. Messerschmitt, “Synchronous data flow,” Proc. IEEE,
vol. 75, no. 9, 1987.
[33] S. Stuijk, “Predictable Mapping of Streaming Applications on Multiprocessors,” Ph.D. dissertation, Eindhoven University of Technology, 2007.
[34] B. Heidergott et. al, Max Plus at Work: Modeling and Analysis of
Synchronized Systems. Princeton University Press, 2006.
[35] http://simit-arm.sourceforge.net/.

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project