null  null
OpenMinTeD
Platform services
distribution
specification
January 10, 2017
Deliverable Code: D6.7
Version: 1.1 – Final
Dissemination level: Public
This report describes and summarizes progress until M18, on specifying and optimizing how
registered TDM services will run and scale on a cloud environment.
H2020-EINFRA-2014-2015 / H2020-EINFRA-2014-2
Topic: EINFRA-1-2014
Managing, preserving and computing with big research data
Research & Innovation action
Grant Agreement 654021
OpenMinTeD Platform services distribution specification
•••
Document Description
D6.7 – OpenMinTeD Platform services distribution specification
WP6 - Platform Design and Implementation
WP participating organizations: ARC, UNIVERSITY OF MANCHESTER, UKP-TUDA, INRA, OU, CNIO,
USFD, GRNET
Contractual Delivery Date: 11/2016
Actual Delivery Date: 1/2017
Nature: Report
Version: 1.0
Public Deliverable
Preparation slip
Name
Organization
Date
From Byron Georgantopoulos,
Vangelis Floros, Stavros
Sachtouris
GRNET
16/11/2016
Edited by Byron Georgantopoulos
GRNET
10/1/2017
ARC
2/1/2017
OU
7/1/2017
ARC
10/1/2017
ARC
10/1/2017
Reviewed by Dimitris Galanis
Petr Knoth
Approved by Natalia Manola
For delivery Mike Chatzopoulos
Document change record
Issue Item
Reason for Change
Author
Organization
V0.1
Draft version
ToC review and finalization
Byron
Georgantopoulos,
Evangelos Floros
GRNET
V0.2
Draft version
Hadoop ecosystem
Byron
Georgantopoulos,
GRNET
V0.3
Draft version
Galaxy and Docker integration
Byron
GRNET
Public
Page 1 of 40
OpenMinTeD Platform services distribution specification
•••
Georgantopoulos,
Evangelos Floros,
Stavros Sachtouris
V0.4
Draft version
Final version ready for internal WP6 Byron
review
Georgantopoulos
GRNET
V0.5
Draft version
Internal WP6 review comments
Dimitris Galanis
ARC
Petr Knoth
OU
V1.0
Final version
Incorporating reviewers’ comments. Byron
Extend Introduction
Georgantopoulos
GRNET
V1.1
Revised
version
Quality verification
Byron
Georgantopoulos
GRNET
Public
Page 2 of 40
OpenMinTeD Platform services distribution specification
•••
Table of Contents
1.
INTRODUCTION .......................................................................................................................................9
2.
OPENMINTED SERVICES IN A NUTSHELL .................................................................................................11
3.
DEPLOYING OMTD WORKFLOW SERVICES .............................................................................................. 13
3.2.1
3.2.2
3.2.3
3.2.4
3.3.1
3.3.2
3.3.3
3.3.4
3.3.5
3.4.1
3.4.2
3.5.1
3.5.2
3.5.3
4.
DISTRIBUTED PROCESSING AND STORAGE WITH HADOOP .....................................................................20
4.2.1
4.3.1
4.3.2
4.3.3
5.
INTRODUCTION .......................................................................................................................................13
USE CASES ............................................................................................................................................. 13
ACCESSING RESULTS .......................................................................................................................................... 13
RUNNING EXISTING WORKFLOWS ........................................................................................................................ 13
SETTING UP TDM COMPONENTS AND WORKFLOWS ............................................................................................... 13
CUSTOMIZED CASES .......................................................................................................................................... 13
SYSTEM OVERVIEW ..................................................................................................................................14
THE PORTAL .................................................................................................................................................... 14
THE REGISTRY .................................................................................................................................................. 14
THE GALAXY WORKFLOW ENGINE ........................................................................................................................ 14
MARATHON/MESOS ......................................................................................................................................... 14
THE PITHOS+ STORAGE BACKEND........................................................................................................................ 15
DOCKERIZING TDM COMPONENTS .............................................................................................................. 15
DOCKER TOOLS AND TERMINOLOGY ..................................................................................................................... 15
DOCKERIZING FOR OPENMINTED ....................................................................................................................... 16
ACCESS CONTROL AND ACCOUNTING ............................................................................................................ 16
INFRASTRUCTURE LAYER .................................................................................................................................... 17
OPENMINTED COMPONENTS LAYER .................................................................................................................... 17
END USER LAYER............................................................................................................................................... 17
OPERATION FLOW AND COMPONENT INTERCONNECTION................................................................................... 18
INTRODUCTION .......................................................................................................................................20
SCALING WITH MAPREDUCE & HADOOP .......................................................................................................20
YARN (HADOOP 2.X) ....................................................................................................................................... 21
HADOOP ON CLOUD INFRASTRUCTURE .......................................................................................................... 22
GENERIC ARCHITECTURE .................................................................................................................................... 22
GRNET CLOUD WITH ~ORKA .............................................................................................................................. 22
PITHOS+ AS HDFS STORAGE BACKEND ................................................................................................................. 23
CONCLUSION .......................................................................................................................................... 23
ANNEX 1: RUN HADOOP ON DOCKER SWARM ....................................................................................... 24
SETUP UP THE SWARM CLUSTER .................................................................................................................. 24
Public
Page 3 of 40
OpenMinTeD Platform services distribution specification
•••
6.
ANNEX 2: RUN HADOOP YARN ON SLIDER.............................................................................................. 29
7.
ANNEX 3: RUN HADOOP YARN WITH GRNET ~ORKA............................................................................... 35
INTRODUCTION .......................................................................................................................................35
INFORMATION ABOUT ~ORKA IMAGES .......................................................................................................... 35
~ORKA OVERVIEW PAGE ............................................................................................................................ 35
CREATING A HADOOP CLUSTER ................................................................................................................... 36
~ORKA CLI EXAMPLES ............................................................................................................................... 38
8.
REFERENCES .......................................................................................................................................... 40
Public
Page 4 of 40
OpenMinTeD Platform services distribution specification
•••
Table of Figures
Figure 1. Overall Architecture of the OpenMinTeD Platform ____________________________________________________ 11
Figure 2: Component of a selected Hadoop image ___________________________________________________________ 35
Figure 3: List of all created Hadoop clusters ________________________________________________________________ 36
Figure 4: Snapshot of ~orka Hadoop creation form __________________________________________________________ 37
Figure 5: List of all jobs in Hadoop cluster _________________________________________________________________ 38
Public
Page 5 of 40
OpenMinTeD Platform services distribution specification
•••
Disclaimer
This document contains description of the OpenMinTeD project findings, work and products. Certain
parts of it might be under partner Intellectual Property Right (IPR) rules so, prior to using its content
please contact the consortium head for approval.
In case you believe that this document harms in any way IPR held by you as a person or as a
representative of an entity, please do notify us immediately.
The authors of this document have taken any available measure in order for its content to be accurate,
consistent and lawful. However, neither the project consortium as a whole nor the individual partners
that implicitly or explicitly participated in the creation and publication of this document hold any sort
of responsibility that might occur as a result of using its content.
This publication has been produced with the assistance of the European Union. The content of this
publication is the sole responsibility of the OpenMinTeD consortium and can in no way be taken to
reflect the views of the European Union.
The European Union is established in accordance with the Treaty
on European Union (Maastricht). There are currently 28 Member
States of the Union. It is based on the European Communities
and the member states cooperation in the fields of Common
Foreign and Security Policy and Justice and Home Affairs. The five
main institutions of the European Union are the European
Parliament, the Council of Ministers, the European Commission,
the Court of Justice and the Court of Auditors.
(http://europa.eu.int/)
OpenMinTeD is a project funded by the European Union (Grant Agreement No 654021).
Public
Page 6 of 40
OpenMinTeD Platform services distribution specification
•••
Acronyms
OMTD
TDM
GRNET
GEANT
AAI
VM
HCFS
HDFS
S3
FTP
UI
CORE
OAI-PMH
IAAS
XML
JSON
REST
URL
RAM
GB
API
Public
OpenMinTeD
Text and Data Mining
Greek Research and Technology Network
Gigabit European Academic Network
Authentication and Authorization Infrastructure
Virtual Machine
Hadoop Compatible File System
Hadoop Distributed File System
Simple Storage System
File Transfer Protocol
User Interface
COnnecting Repositories
Open Archives Initiative – Protocol for Metadata Harvesting
Infrastructure as a Service
Extensible Markup Language
JavaScript Object Notation
REpresentational State Transfer
Uniform Resource Locator
Random Access Memory
Gigabyte
Application Program Interface
Page 7 of 40
OpenMinTeD Platform services distribution specification
•••
Publishable Summary
D6.7 reports and summarizes progress until M18, on optimizing OpenMinTeD service deployment and
execution via scalable VM-based services. The described work falls under Layer 3 of the high-level
architecture: interoperability to shared storage and computing resources. As TDM services require big
data technologies to run in parallel and within distributed cloud environments, effort on WP6 has
focused on exploring, experimenting and outlining the following:
•
•
Public
Deployment of OMTD workflow engine and TDM components (tools & services) that
constitute the workflow, on cloud infrastructures,
Handling scaling requirements, i.e. data size, parallel users, performance, etc.
Page 8 of 40
OpenMinTeD Platform services distribution specification
•••
1. Introduction
OpenMinTeD aspires to create an infrastructure that fosters and facilitates the use of text and data
mining technologies in the scientific publications world and beyond, by both application domain users
and text-mining experts.
The WP6 activity has a set goal to handle the design and development of the OpenMinTeD platform
services. In specific, it will (a) define the conceptual architecture and design specification of the
platform and (b) design and implement the set of services that the platform consists of. The
OpenMinTeD platform follows a distributed and modular approach to its architecture and the
deployment of its core services.
D6.7 report, the first of two reports (the second scheduled on M26),
describes and summarizes
progress until M18, on specifying and optimizing how registered TDM services will run and scale on a
cloud environment. The described work falls under Layer 3 of the overall architecture: interoperability
to shared storage and computing resources. As TDM services require big data technologies to run in
parallel and within distributed cloud environments, effort on WP6 has focused on exploring,
experimenting, assessing and outlining the following:
•
Deployment of OMTD workflow engine and TDM components (tools & services) that
constitute the workflow, on cloud infrastructures,
•
Addressing scaling requirements, i.e. data size, parallel users, performance, etc.
To this end, choice of workflow software proved very crucial, since it effectively serves as the
connecting layer between the Registry and the actual execution of TDM services. The final decision
over Galaxy 1 was also influenced by existing and well documented Docker 2 integration work.
Arguments in favor of Galaxy were mainly: open source with community support, active community
which provides examples, help and even ready-to-use components, tested in cases very similar to
OpenMinTeD, adequate support for a variety of popular clouds, storage back-ends and similar
technologies, including Docker. Containerization through Docker will provide a method of defining and
running TDM components (tools & services) in a totally isolated, idempotent manner. Dockerizing the
components is required for registering them in the OpenMinTeD workflow ecosystem, since this solves
various interoperability issues that would otherwise emerge as consequence of using different OS,
programming languages, libraries, dependencies, etc. under a single platform.
Concerning scalability, a plethora of open source solutions exist and several have already been tried
out, particularly on two fronts: Docker-related schedulers (Docker Swarm, Apache Mesos and the
container orchestration platform Apache Marathon) and Hadoop-related ones (YARN). In specific:
1 https://galaxyproject.org/
2 https://www.docker.com/
Public
Page 9 of 40
OpenMinTeD Platform services distribution specification
•••
•
•
•
•
Public
we run some experiments on various tools (i.e., YARN with Slider, Mesos/Marathon, Docker
Swarm) on ~okeanos IaaS
we have created re-usable know-how (setup guides) on how to install and deploy these services
we have a tested solution for running and deploying Hadoop clusters on ~okeanos
work on storage backend support has also started, towards developing an API connector
between Galaxy and GRNET Pithos+ cloud object storage system.
Page 10 of 40
OpenMinTeD Platform services distribution specification
•••
2. OpenMinTeD Services in a Nutshell
The following figure illustrates the high-level architecture of the OpenMinTeD platform [1]:
Figure 1. Overall Architecture of the OpenMinTeD Platform
OMTD will offer a Registry service for storing, browsing, downloading, searching and managing various
resources such as publications, processing components (e.g. a named entity tagger or sentence
splitter) and languages resources (e.g. machine learning models, lexica, thesauri). These resources will
be registered and, concerning publications, externally linked in OMTD by exploiting harvesting
capabilities of OpenAIRE and CORE, (e.g. OAI-PMH, FTP, etc.), specified as Docker files for containerized
execution, and documented with high quality metadata. The Workflow Editor service of the platform
will guide users via an appropriate UI (based on the respective metadata) in creating interoperable
workflows of TDM components; the workflows will be executed from the Workflow service in a cloud
infrastructure or locally, using the above Dockerfiles. OMTD users will be able to annotate the
publications (texts) using the Annotation Editor service and create datasets which can be used in
workflows; e.g. for evaluation purposes. Services under the OMTD Enabling Layer will facilitate
monitoring, automatic deployment and re-scaling of the platform. Also, appropriate Authentication
and Authorization Infrastructure (AAI) services will support the platform. All the aforementioned
Public
Page 11 of 40
OpenMinTeD Platform services distribution specification
•••
services will communicate via appropriate REST-based APIs and will be integrated and offered in a
uniform way via the web Portal of OMTD.
Public
Page 12 of 40
OpenMinTeD Platform services distribution specification
•••
3. Deploying OMTD workflow services
Introduction
In this Section, we explore how all OpenMinTeD services will be tuned to execute dockerized
workflows on Galaxy. After presenting the basic components, we define the operation flow of the
system. We briefly describe the process of dockerizing TDM tools and services and clarify the
interconnection between Galaxy, the Registry service and the Pithos+ storage backend. We will use a
few basic use cases as guides to help demonstrate how the system will operate from the point of view
of an administrator, a task or workflow developer, and the OpenMinTeD users.
Cloud computing resources to support the above activities are provided for the purposes of the project
by GRNET. GRNET currently operates a Data Center in Athens, Greece, comprising a large number of
physical servers, storage servers and the necessary network infrastructure that allows high-speed
interconnectivity with the pan-European research network GÉANT. GRNET has developed its own inhouse cloud computing solution named ~okeanos and has been offering IaaS cloud services for free to
the Greek Research and Academic community (primarily) and to European researchers in the context
of relevant EC projects.
More details about ~okeanos and its integration with Galaxy and container technologies focusing on
the Docker tool ecosystem are documented on D8.1 Infrastructure report [2].
Use cases
3.2.1 Accessing results
The following use cases, in order of simplicity, are indicative of the complexity in terms of services
interoperability; the first requires the least operations and interconnections, the last involves most of
the infrastructure and supporting services of the OMTD system.
3.2.2 Running existing workflows
In this use case, we assume users are authenticated and authorized to use the system. They find an
already registered and installed workflow and run it on OMTD datasets, i.e. a set of publications.
3.2.3 Setting up TDM components and workflows
We assume users possess the authority and skills to create their own workflows. They are interested
not only on running them, but also on sharing them with selected colleagues or the community.
3.2.4 Customized cases
The system may allow using datasets stored on custom and external locations, storing and exporting
intermediate results and plugging additional functionalities on specific workflows.
Public
Page 13 of 40
OpenMinTeD Platform services distribution specification
•••
System Overview
3.3.1 The Portal
The OMTD Portal acts as the front end and the UI of the system, guiding users to authenticate and
authorize themselves. Authentication and Authorization is delegated to the AAI service and, if
successful, the control returns to the Portal, updated with the actions permitted for this particular
user.
3.3.2 The Registry
The Registry service maintains a record of TDM resources and data sources and acts as a coordinator
for data source handling and workflow management. It also manages access control to related services
(e.g., workflow editor, data source repositories), validates text and data mining resources and handles
dataset storage and sharing issues. The Registry features a UI to let users look up and reuse resources,
data sources and workflows (e.g., run an existing workflow on Galaxy). Most of these operations (e.g.,
access control, workflow execution, data storage) are delegated to other components in a transparent
manner.
3.3.3 The Galaxy workflow engine
Galaxy is a tested, well-known workflow engine supported by a growing community. It will be placed at
the heart of the OpenMinTeD system and it will be responsible for running dockerized workflows on
corpora stored on the Pithos+3 storage backend. Pithos+ will also accept and store the result data (e.g.
annotations). Users will be able to access Galaxy and watch the progress of their workflows, track
possible errors and register their own components.
In order to achieve access control, there must be a mapping between Registry authorized and Galaxy
users. From the Galaxy point of view, access control is handled by the Registry service.
Galaxy can read from and store to external storage back-ends (e.g. Amazon S34) or at the local storage
(e.g., a block storage device attached on the Galaxy VM).
3.3.4 Marathon/Mesos
Each workflow consists of a series of individual TDM components, which must be somehow executed.
The workflow engine will delegate this operation to Marathon5 and Mesos6. These tools can provision
virtual machines on the cloud and run tasks (e.g., as Docker containers). They feature load balancing,
extensibility, adequate monitoring and logging. This subsystem will serve as an intermediate layer
between Galaxy and the cloud infrastructure.
3 https://pithos.okeanos.grnet.gr
4 http://galaxy.readthedocs.io/en/stable/lib/galaxy.objectstore.html
5 https://mesosphere.github.io/marathon/
6 http://mesos.apache.org/
Public
Page 14 of 40
OpenMinTeD Platform services distribution specification
•••
3.3.5 The Pithos+ Storage backend
Pithos+ is the object storage service of GRNET ~okeanos 7 infrastructure; it can be accessed from
external applications through the Pithos REST API part of the kamaki python library 8. Kamaki library is
an open source, multipurpose, interactive CLI tool and client development API for managing Synnefo as
well as other OpenStack clouds.
In the context of OpenMinTeD, there are currently two implementation options available for input
corpora and output results: Galaxy-accessible block device (e.g., a volume attached on the VM running
Galaxy) and Pithos+ (cloud storage backend). The system will make use of the sharing capabilities of
Pithos+ to provide result publication and implement access restrictions when needed.
Dockerizing TDM components
In the context of OpenMinTeD, workflows are composed of simple autonomous components (TDM
tools) chained together (e.g., Tokenization -> Sentence Splitting -> POS Tagging -> Lemmatization ->
Named Entity Recognition). OMTD task developers must conform to certain guidelines and rules in
order to ensure compatibility (i.e. interoperability) with other components, reusability in versatile
contexts and reduced maintenance complexity. Last, but not least, components must be able to be
spawned by the workflow engine. These goals are achieved by bundling each component (or a set of
components) as a Docker container. Developers working with Docker commonly refer to this process
as “dockerization”.
3.4.1 Docker tools and terminology
Docker is a collection of tools for isolating software subsystems by utilizing the container capabilities of
(primarily) the Linux kernel and other operating systems. Docker images and containers can be
developed and run on all popular Linux distributions, Mac OS X and Windows. More details and
examples can be found in the Docker documentation pages 9.
A Docker image is a recipe for building isolated containers. It is built from a Dockerfile which is
typically located at the root of the code tree, maintained by the task developers and contains
instructions to the Docker engine e.g., what base Linux distribution should be used, software
requirements to be installed, code repositories, compiling instructions, networking settings,
environment variables etc. More information can be found in the official Dockerfile reference 10.
Typically, developers execute “docker build” to build Docker images from a Dockerfile and release
them on an appropriate repository (registry). Docker Hub11 provides a free publicly available Docker
7 https://okeanos.grnet.gr
8 https://www.synnefo.org/docs/kamaki/latest/developers/clients-api.html
9 https://docs.docker.com/
10 https://docs.docker.com/engine/reference/builder/
11 https://hub.docker.com/
Public
Page 15 of 40
OpenMinTeD Platform services distribution specification
•••
repository but since the respective software is open source, a similar repository can be setup internally
(private registry) to serve the needs of OpenMinTeD. It features archiving, version tags and searching.
In order to use a Docker image, users must execute the “docker run” command, which creates an
instance of the Docker image. These instances are called Docker containers and can be started,
stopped, snapshotted and updated. When a Docker container is started successfully, the application is
up and running. Multiple containers can be created and run simultaneously from the same image, even
on the same VM. Typically, a TDM task software wrapped as a Docker image should contain
instructions for starting and parameterizing a Docker container.
3.4.2 Dockerizing for OpenMinTeD
A workflow engine would spawn (“docker run”) a container with predefined options that ensure data
chaining between components. Since the output of a component is the input of its successor, the
workflow engine should be able to let each component container know where the input and the
output should be placed, using Docker volumes and, optionally, environment variables.
A Docker volume is a mapping between a location in the host file system and the container file system,
offering persistent storage. It can be seen as a way to bi-directionally exchange data between the
component and the workflow engine.
For instance, to execute task N in a succession of tasks, the workflow engine may execute a command
like this (mounting a host directory as a data volume for the container using the “:” designator):
$ docker run \
-v /workflow/datasets/input-for-task-N:/container/input \
-v /workflow/datasets/input-for-task-N-plus-1:/container/output \
task-N-image:latest
When the task has completed its execution, the engine will proceed to the next task (N plus 1) in the
workflow:
$ docker run \
-v /workflow/datasets/input-for-task-N-plus-1:/container/input \
-v /workflow/datasets/input-for-task-N-plus-2:/container/output \
task-N-plus-1-image:latest
Access control and accounting
There are three types of roles in the OpenMinTeD system:
(1) Starting from the bottom, virtual resources are assigned and quoted to infrastructure users,
who deploy and maintain the system components.
Public
Page 16 of 40
OpenMinTeD Platform services distribution specification
•••
(2) Each component may feature its own role management, which should be kept as simple as
possible.
(3) At the top, the Registry implements the end user access management by utilizing the features
of each component and by relying on a separate authentication and authorization component.
3.5.1 Infrastructure layer
A core design choice regarding cloud infrastructure is to deploy all services (Portal, Registry, Workflow
Editor & Execution, etc.) in one single cloud account. More than one account can be created if such a
requirement is inferred by the system design, but even in that case, cloud accounts will not correspond
to end user accounts. However, running all services in one account reduces deployment and
maintenance complexity without loss of performance or availability. Infrastructure accounting, statistics
and quotas are all managed in the context of this single account. One OpenMinTeD account will keep
the ownership of all infrastructure: virtual machines running the system components, virtual machines
provisioned to host Docker containers, Pithos+ storage space, public IPs and inter-component
networking infrastructure. Logs and statistics related to the infrastructure utilization are kept and
watched by this single user.
3.5.2 OpenMinTeD components layer
The core components of the platform (see Figure 1. Overall Architecture of the OpenMinTeD Platform)
may define their own user roles. In specific, the Registry and the Galaxy components will support their
own kinds of user roles which can be mapped/linked (if needed) to the user roles defined and
supported by the AAI service. For instance, a user with read-only access rights (e.g., can only have
access to results shared with them) will not be assigned a Galaxy account, while a workflow owner
must have access to either an atomic or a shared Galaxy account, automatically handled by the system.
These accounts may be prepared in advance (e.g., a pool of Galaxy accounts) or created on demand to
serve each individual need.
3.5.3 End user layer
The entry point for each user is the portal front-end application. Authentication and authorization is
delegated to the AAI component. The AAI component should maintain a directory of roles and access
rights for each user. Actual resources will be hidden from the end user i.e., the virtual machines running
the Docker containers on which user tasks are executed, or the Pithos+ storage location where the user
uploads datasets or from which they download results.
Access to OpenMinTeD components may also be transparent, depending on the user role. The details
of how and where a user’s workflow is executed can be hidden from them, but for some users the
system may authorize direct workflow engine access (Galaxy as a Service).
Public
Page 17 of 40
OpenMinTeD Platform services distribution specification
•••
Operation flow and component interconnection
In this subsection, we will describe how components interconnect, using some general use cases as
guides and examples.
The entry point for every user is the portal front-end. User authentication procedure happens there,
which is typically delegated to the AAI component. The latter can authenticate federated users and
respond back to the portal, providing useful authorization information as well.
The authorization process is handled by the Registry. A mutable user directory (e.g., an LDAP system)
should be maintained (either in the Registry or as a separate component), in order to restrict or provide
access to datasets, results and workflows. Based on this information, the Registry component can
decide if the users can upload datasets, run workflows or have access to already committed jobs.
In use case 3.2.1 (Accessing results), the Registry must provide a searching mechanism. Each piece of
data (sources, results) is stored in Pithos+ and has a public URL assigned to it. All public URLs are
preserved in the Registry database as well. The Registry service checks whether the user is authorized
for this piece of data, and if they are, the corresponding Pithos+ public link is provided.
In use case 3.2.2 (Running existing workflows), the user searches for predefined workflows and
attempts to run them on custom datasets (corpora of publications) stored on an external system (e.g.,
the user’s local disk or a storage location on another cloud). As pointed in Section 2, in most cases this
should be done by creating and gathering the datasets from OpenAIRE/CORE directly within the OMTD
platform. An authorization step follows and, if successful, the user can provide the dataset through e.g.
an upload web form. The Registry uses the Pithos+ API to upload this dataset to Pithos+, in a Pithos+
path hidden from the user.
The Registry communicates with Galaxy through the Galaxy API and requests access to a Galaxy
account which owns the corresponding workflow. Registry requests Galaxy to execute the workflow,
thus passing control to the latter. Galaxy pulls the user dataset from Pithos+ using a kamaki12 backend
and then delegates the execution of the first container to Mesos/Marathon. The latter decides on
which virtual machine this container should run on, returning a handle to Galaxy. In the case of two
successive components, when the first is completed, Mesos/Marathon is called to run the next one,
and so on until the whole workflow is consumed. Galaxy can be configured to run components in
parallel as well. The results are pushed to Pithos+ through the same kamaki backend and a new Pithos+
location is returned to Registry along with the control of the flow.
Finally, a link to the results is revealed to the user, who can proceed to download and use them.
In use case 3.2.3 (Setting up TDM components and workflows), the users are provided with guidelines
on how to dockerize TDM components for OpenMinTeD. This information will be publicly accessible, so
12 https://www.synnefo.org/docs/kamaki/latest
Public
Page 18 of 40
OpenMinTeD Platform services distribution specification
•••
that users can compose their own OMTD-compliant components. When they decide to register them
and after they pass through the authorization check, the Registry picks and pushes their components to
the Docker private repository. Although the components are internally available through the Docker
repository, an auditing procedure (a human managed moderation step) may be required before the
Registry marks them as ready for the workflow engine.
In order to create a workflow, the Registry communicates with Galaxy through the Galaxy API and
requests access to a Galaxy account, through which the user is redirected to the Galaxy UI. Now, TDM
components can be selected, removed, arranged and connected in a graphic manner. Each TDM
component corresponds to a Docker image in the internal (private) Docker repository. The information
on the available components is maintained in the Registry, which controls how it will be provided to
Galaxy, typically with a structured file containing a URL to the component. Since each TDM component
is a Docker image, the URL is the image location, so that Galaxy can pull it directly from the repository.
In the next step, control passes to Galaxy. The user is given access to the Galaxy graphic editor, where
they can compose their workflows by chaining individual components. When they are done, the control
returns to the Registry where they are required to provide metadata (e.g. workflow descriptions).
Running the workflow is already covered in use case 3.2.2.
In use case 3.2.4 (Customized cases), a user needs access to intermediate results or custom Galaxy
features (e.g. errors). Assuming they are authorized, they can deal directly with Galaxy.
Special Docker images can be built with appropriate software and configurations in order to store
intermediate results at a place where the user has access, or extract meaningful information by
processing intermediate data. Component developers can build and chain custom Docker images of
this nature or use the ones provides by OpenMinTeD. A typical intermediate image would, for instance,
copy the incoming data to a cloud location accessible by the user and then return to Galaxy.
Public
Page 19 of 40
OpenMinTeD Platform services distribution specification
•••
4. Distributed processing and storage with Hadoop
Introduction
Besides the above described Galaxy workflow engine system, the project has stated explicitly its
interest on big data technologies, especially Hadoop, for delivering and scaling TDM components.
One option for this is to describe, register, and make Hadoop-ized TDM components searchable in the
OMTD Registry, the latter could then launch these components in an already existing multi-node
Hadoop cluster (or launch them in a temporary one). TDM components can be expressed in the
MapReduce paradigm and ported to Hadoop with minimal effort, since various tools like tokenization,
splitting, lemmatizing, etc. are perfectly fit for the shared-nothing architecture of Hadoop.
Another option is to use the Docker container executor13 which basically launches dockerized TDM
components as Hadoop 2.x (YARN) containers.
A third option is to use Hadoop-based programs in Galaxy workflows (attempts are documented in
https://github.com/crs4/hadoop-galaxy
and
https://wiki.galaxyproject.org/CloudMan/Hadoop).
Apache Mesos also provides a Hadoop environment in https://github.com/mesos/hadoop.
In the rest of this section, we provide an overview of the Hadoop ecosystem. Running under Hadoop
can benefit from the ready software stack to run and distribute services at a large number of nodes,
compose Hadoop jobs into workflows with Oozie14, as well boost speed execution with in-memory
processing through frameworks such as Spark 15 and Flink16. We further describe some associated
experiments in the annexes 5, 6, 7.
Scaling with MapReduce & Hadoop
MapReduce is the most popular programming model for big data processing. Each algorithmic problem
and its associated data are sliced in small chunks and assigned for execution to each cluster node (Map
phase). Afterwards, results from each node are synthesized to compute the final solution (Reduce
phase). Basic principles of this model are the transfer and execute of code to the node that owns the
data chunk (“move code to the data”) and the linear scanning of data, from start to end (“batch
mode”).
Apache Hadoop17 platform is the most widely used implementation of the above model. It is open
source, general application software, and does not require expensive hardware or storage, in contrast
13 https://hadoop.apache.org/docs/r2.7.3/hadoop-yarn/hadoop-yarn-site/DockerContainerExecutor.html
14 https://oozie.apache.org
15 http://spark.apache.org/
16 http://flink.apache.org/
17 http://hadoop.apache.org/
Public
Page 20 of 40
OpenMinTeD Platform services distribution specification
•••
it can horizontally scale out with commodity hardware. It allows users to develop and execute
distributed applications on very large datasets without worrying about the details and technicalities of
parallelization, data replication, and fault-tolerance.
In a typical Hadoop setup, a master node is responsible for managing both the data (datanode VMs)
and processing (tasktracker VMs) available resources, called “slave” nodes. HDFS is the distributed data
storage system of Hadoop. In general, files in Hadoop are split into chunks of configurable size (usually
64MB or 128MB) and each chunk is stored in a datanode, i.e. a VM that stores data. The namenode
daemon in the master VM holds a metadata table that contains the necessary information to construct
a big data file from its component chunks. HDFS also stores each chunk in more than one datanodes
(replication factor, again configurable) in order to achieve fault-tolerance when a datanode becomes
unavailable. HDFS has some but not all capabilities of a full filesystem. It is generally designed towards
few and big data files that are processed from start to end, and cannot be appended.
Around the Hadoop system (called Hadoop commons) which basically consists of HDFS and the
execution mechanism for running MapReduce jobs has evolved a full ecosystem to cover needs not (or
not easily) expressed with the original MapReduce platform. This multitude of software and
frameworks covers a wide range of applications:
•
•
•
•
•
•
Workflow engines on MapReduce jobs, such as Oozie,
NoSQL databases and columnar databases such as Hive and HBase,
Machine Learning libraries and toolsets such as Mahout and MLib,
Data ingestion architectures such as Kafka and Flume,
Lambda architectures to handle both batch and real-time data such as Spark and Flink,
Complete solutions, with rich GUI and a plethora of pre-installed components, either free or
commercial traits such as Cloudera, Hortonworks and MapR.
4.2.1 YARN (Hadoop 2.x)
The next generation (Hadoop 2.x) of the original Apache Hadoop platform is YARN (Yet Another
Resource Negotiator). YARN is a resource management platform aiming primarily at solving the scaling
issues MapReduce faced when deploying clusters with a huge number of nodes (in the range of
thousand nodes). Two separate daemons have the responsibility to (i) handle the cluster nodes (Node
Manager) and (ii) manage each application’s lifecycle (Application Master). The latter negotiates
cluster resources with the former – resources are expressed with the term Container; each Container
has an upper memory limit – and then executes the application processes inside the Containers. In this
sense, YARN is more abstract than MapReduce, in fact a MapReduce job is just one type of jobs that
can be executed inside a YARN cluster. In summary, the YARN architecture improves the performance
of Hadoop applications and provides support for a wider range of applications besides MapReduce,
and is often called as “the Hadoop OS” since it implements process management and monitoring of
different clients in a multi-tenant environment, security policies, and high-availability.
Public
Page 21 of 40
OpenMinTeD Platform services distribution specification
•••
Hadoop on cloud infrastructure
4.3.1 Generic architecture
Provisioning a YARN cluster in a cloud setting means basically provisioning and configuring VMs with
the Hadoop (or any other ecosystem) software on a number of VMs inside a cloud infrastructure.
Alternatively, one can use automated distributions from Hadoop ecosystem vendors that provision,
install and configure a fully functional YARN cluster. These solutions (some free and commercial) are
provided by:
•
•
•
•
Hortonworks Data Platform, http://hortonworks.com/downloads/#data-platform
Cloudera Manager, https://www.cloudera.com/products/cloudera-manager.html
MapR, https://www.mapr.com/get-started-with-mapr
Apache Ambari, https://ambari.apache.org
And the following are commercial offerings:
•
•
Amazon EMR, https://aws.amazon.com/emr/
Google Cloud, https://cloud.google.com/hadoop/
4.3.2 GRNET cloud with ~orka
~orka18 is a cloud-based platform for deploying and running big data analytics, using the Hadoop
ecosystem. It is designed and developed in the context of GRNET e-Science project19, and provided to
the members of Research and Academic Community as an ~okeanos-powered service. The main goal is
to remove the complexity for researchers to manually provision, deploy and manage multi-node
Hadoop clusters.
With ~orka, big data deployments over ~okeanos infrastructure are easily configured, spawned and
maintained. There is a variety of mature, well established tools and solutions centered around Hadoop:
•
•
•
•
•
Create YARN clusters with selected ~okeanos flavors (hardware specifications) for master and
slave nodes
Choose from a number of pre-cooked images:
o Hadoop Core
o Cloudera
o Hue (Hue User Interface)
o and a consolidated Ecosystem image with: Pig, Hive, HBase, Oozie, Spark, Flume
Run workflows with Oozie (a workflow scheduler system for managing Apache Hadoop jobs)
Stream external data into HDFS with Flume
Manage the Hadoop cluster (Start, Stop, Format, Destroy)
18 https://okeanos.grnet.gr/platforms/orka/
19 https://github.com/grnet/e-science/
Public
Page 22 of 40
OpenMinTeD Platform services distribution specification
•••
Scale the Hadoop cluster (Add, Remove nodes)
• Use Pithos+ as a storage backend, run MapReduce algorithms with input and output directly
from/to Pithos+ files
• Export the Hadoop cluster metadata and algorithmic scripts, so that the algorithm can be
replayed later, in the same or a new cluster, with the same or different parameters
• Create own pre-cooked images, and add them to the ~orka dropdown list of available images
~orka services are available through a Personal Orka Server's dashboard (GUI) as well as a command
line interface (CLI). Both access types are based on a REST-API, that is also provided to the end users.
•
4.3.3 Pithos+ as HDFS storage backend
By default, HDFS uses the native disks of the namenodes to store and retrieve data. Connecting HDFS
to Pithos+ storage backend provides a range of advantages for the ~okeanos ecosystem:
•
OMTD users can access a larger storage, based on their quota, that can surpass the available
disk space on the namenode VMs
Exploit native Pithos+ de-duplication; Many corpora may have large parts matching with preexisting corpora. Thanks to de-duplication, we can save on storage space and upload time
Easier sharing and moving of big files between ~okeanos users and/or other Hadoop clusters
•
•
In order to bring these benefits to ~okeanos users, ~orka has implemented a HDFSPithos+
connector20, similar to the one natively supplied to Hadoop users for the Amazon S3 filesystem21. The
connector effectively creates a Hadoop Compatible File System (HCFS), with a “pithos://” namespace
which allows to list files, copy files between Pithos+ and namenode storage and provide Pithos+ files to
MapReduce jobs. All these commands are also available out-of-the-box from the ~orka CLI.
Conclusion
During the first 18 months of WP6 work, the consortium has thoroughly explored options and finalized
a plan to offer TDM components as Docker images and chain them with the Galaxy workflow software
at a scalable multi-VM environment with Swarm and Mesos/Marathon. In the following period, the
Dockerization of TDM components has to be geared up, so as the OMTD Portal/Registry can offer
complete end-to-end workflows running in the cloud. Storage should also have to be taken care of,
with Pithos+ being a fully functional storage backend for Galaxy workflows. Finally, we have to arrive at
what scale and how Hadoop jobs can be integrated into Galaxy.
20 https://github.com/grnet/e-science/tree/master/pithosfs
21 https://wiki.apache.org/hadoop/AmazonS3/
Public
Page 23 of 40
OpenMinTeD Platform services distribution specification
•••
5. Annex 1: Run Hadoop on Docker Swarm
Docker Swarm22 offers native clustering for Docker. It turns a pool of Docker hosts into a single, virtual
Docker host. Because Docker Swarm serves the standard Docker API, any tool that already
communicates with a Docker daemon, as it happens with Galaxy and dockerized TDM services in
OMTD context, can use Swarm to transparently scale to multiple hosts.
The following is a step-by-step guide to deploy a non-trivial application over a cluster of Synnefo VMs
running interconnected Docker containers. The non-trivial application of choice is Hadoop, because it is
distributed and provisioning, monitoring, scaling and accounting are meaningful requirements.
Setup up the Swarm cluster
The Docker tools we use are:
•
•
•
•
•
docker-engine: the basic docker application for building images and running containers
docker-machine: a tool for provisioning hosts in order to be docker-ready, as well as to prepare
containers for participating in docker clusters
consul: a service-discovery application
docker-swarm: container cluster manager
docker-UCP: a higher-level tool for managing the cluster, comes with a UI
In this example, we have used 4 VMs:
•
•
•
•
the "admin" VM is loaded with tools for the cluster administrator. In a real-life scenario, it could
be the laptop or the PC of the admin.
A VM for consul.
A VM for the swarm master. It can also run some Docker containers, if the swarm master
decides it automatically.
A VM for running Docker containers. Adding more VMs like this is equivalent to scaling up the
cluster resources.
The image id used corresponds to an Ubuntu server LTS image, while the flavour corresponds to a 4
core, 4MB RAM, 20 GB VM. These requirements were sufficient for our tests.
Setup the admin machine
In this example, we are going to create an admin VM, but you can use your local host if instead. We are
naming the admin VM "docker 0":
22 https://www.docker.com/products/docker-swarm , https://docs.docker.com/swarm/overview/
Public
Page 24 of 40
OpenMinTeD Platform services distribution specification
•••
$ kamaki server create --name "docker 0" --image-id fbfefd37-755a-4746-94ce-8d6d75c279f6 --flavor-id 25 -p
/home/user/.ssh/id_rsa.pub,/home/user/.ssh/authorized_keys,user,user,0600 --project-id a6f26140-0af2-4e529984-51a1df051619 -w
Login with ssh and update, upgrade:
$ sudo apt update -y && sudo apt upgrade -y
Install docker-engine:
$ sudo apt-get install apt-transport-https ca-certificates
$ sudo apt-key adv --keyserver hkp://p80.pool.sks-keyservers.net:80 --recv-keys
58118E89F3A912897C070ADBF76221572C52609D
$ echo deb https://apt.dockerproject.org/repo ubuntu-trusty main > /etc/apt/sources.list.d/docker.list # for Ubuntu
Trusty
$ sudo apt update
$ sudo apt-get purge lxc-docker
$ apt-cache policy docker-engine
$ sudo apt-get install docker-engine
$ sudo service docker start
Install docker-machine and as root:
# curl -L https://github.com/docker/machine/releases/download/v0.7.0/docker-machine-`uname -s`-`uname -m` >
/usr/local/bin/docker-machine && \
chmod +x /usr/local/bin/docker-machine
Also, generate a pair of keys to be used for provisioning and managing the nodes of the cluster::
$ ssh-keygen -t rsa -b 4096
(passwordless)
Install/setup kamaki:
$ sudo apt-get install python-software-properties
$ sudo add-apt-repository ppa:grnet/synnefo
$ sudo apt-get update
$ sudo apt-get install kamaki
$ kamaki config set cloud.~okeanos.url <Synnefo authentication URL>
$ kamaki config set cloud.~okeanos.token <Synnefo user token>
Public
Page 25 of 40
OpenMinTeD Platform services distribution specification
•••
Create 3 Synnefo VMs
From the admin machine, create three VMs. Make sure to allow PKI root access::
$ kamaki server create --name "docker " \
--image-id fbfefd37-755a-4746-94ce-8d6d75c279f6 --flavor-id 25 \
-p /home/user/.ssh/id_rsa.pub,/root/.ssh/authorized_keys,root,root,0600 \
--project-id a6f26140-0af2-4e52-9984-51a1df051619 -w --cluster-size 3
In this example, we created three identical VMs for simplicity, using the "--cluster-size" option, but you
can create each VM separately if needed.
Setup consul (service discovery)
Docker can use many different service discovery systems, but this is the one used on their guides. We
don't currently have strong preference for Consul, except that it seems to work with docker tools.
Provision this VM using docker-machine:
$ docker-machine create --driver generic --generic-ip-address <docker 1 IP>\
--generic-ssh-key /home/user/.ssh/id_rsa consul
Now, setup the environment on admin machine to manage the (remote) consul container:
$ eval $(docker-machine env consul)
$ docker run -d -p "8500:8500" -h "consul" progrium/consul -server -bootstrap
Setup swarm node
Swarm master will run on VM docker 2. The procedure is now familiar:
$ docker-machine create --driver generic --generic-ip-address <docker 2 IP>\
--generic-ssh-key /home/user/.ssh/id_rsa --swarm --swarm-master\
--swarm-discovery="consul://$(docker-machine ip consul):8500"\
--engine-opt="cluster-store=consul://$(docker-machine ip consul):8500"\
--engine-opt="cluster-advertise=eth1:2376" swarm-master
Provision a swarm node
We already have a spare VM (docker 3), so we don't need to:
$ docker-machine create --driver generic --generic-ip-address <docker 3 IP>\
--generic-ssh-key /home/user/.ssh/id_rsa --swarm \
--swarm-discovery="consul://$(docker-machine ip consul):8500"\
Public
Page 26 of 40
OpenMinTeD Platform services distribution specification
•••
--engine-opt="cluster-store=consul://$(docker-machine ip consul):8500"\
--engine-opt="cluster-advertise=eth1:2376" swarm-node
Setup overlay network
Create an overlay network. Don't forget to "eval" your swarm-master environment first:
$ eval $(docker-machine env --swarm swarm-master)
$ docker network create --driver overlay cluster-network
Setup docker UCP
Typically, we can setup an overlay network to complete the cluster setup. This will be done
automatically by the UCP service.
Setup UCP on docker 2 (the swarm-master VM):
$ docker-machine ssh swarm-master
$ docker run --rm -it -v /var/run/docker.sock:/var/run/docker.sock \
--name ucp docker/ucp install -i --swarm-port 3367 \
--host-address <swarm-master IP>
(you will need to set a password for UCP admin)
Check your UCP at https://<docker 2 IP>:3367
UCP requires a license e.g. to add new nodes to the cluster. There is a free trial license for testing.
Run Hadoop on the cluster
To deploy a Hadoop system on our swarm cluster, we need to be provided with a Hadoop docker
image e.g., "kiwenlau/hadoop:1.0".
We will set up a hadoop master with two slaves:
$ docker -H $(docker-machine ip swarm-master):3376 run -itd --net cluster-network -p 50070:50070 -p 8088:8088
--name hadoop-master --hostname hadoop-master kiwenlau/hadoop:1.0
$ docker -H $(docker-machine ip swarm-master):3376 run -itd --net cluster-network --name hadoop-slave1 -hostname hadoop-slave1 kiwenlau/hadoop:1.0
$ docker -H $(docker-machine ip swarm-master):3376 run -itd --net cluster-network --name hadoop-slave2 -hostname hadoop-slave2 kiwenlau/hadoop:1.0
Each Hadoop element runs on a container. The location of each container is managed by the swarmmaster. We can use UCP to check were everything runs, or even from the command line.
Check what containers run on what VM:
$ docker -H(docker-machine ip swarm-master):3376 ps
Public
Page 27 of 40
OpenMinTeD Platform services distribution specification
•••
$ docker -H(docker-machine ip swarm-node):3376 ps
In some of our examples, the Hadoop master was installed on the swarm-node, while one or two of
the slaves on the swarm-master. Still, Hadoop admins do not need this information. They can manage
their Hadoop containers as if they run on their local system.
Let's start the Hadoop cluster and run a sample WordCount job on it:
$ docker exec -it hadoop-master bash
# ./start-hadoop.sh
# ./run-wordcount.sh
Hadoop offers UI access for the cluster on http://<hadoop master IP>:8088/cluster and HDFS UI
access on http://<hadoop master IP>:50070.
Find the Hadoop master IP with simple commands like:
$ docker exec -it hadoop-master ifconfig
Public
Page 28 of 40
OpenMinTeD Platform services distribution specification
•••
6. Annex 2: Run Hadoop YARN on Slider
YARN, being primarily the virtual cluster manager and resource scheduler of Apache Hadoop, can be
used for other distributed applications as well. This is achieved by configuring YARN to handle e.g.
Docker containers and also by providing a custom scheduler to serve the application needs. The
process is greatly simplified with Apache Slider23, which automates the deployment of distributed
applications (with or without Docker) on an existing YARN cluster. Scaling, monitoring and application
handling are also improved with Slider, either from the “servers” command line interface or through
the Slider REST API.
Setup
The following steps are required to successfully run YARN jobs with Slider.
•
•
•
•
•
Create 3 VMs with kamaki
Setup Hadoop master
Setup a Hadoop cluster
Install slider
Run docker apps with slider
Provisioning
Create 3 VMs with kamaki, with the Hadoop images
kamaki server create --name "hadoop " --cluster-size 3 --image-id be029a82-d61b-4d44-9bb3-ec2c2911e631 -flavor-id 18 --project-id a6f26140-0af2-4e52-9984-51a1df051619 -p
/home/user/.ssh/id_rsa.pub,/root/.ssh/authorized_keys,root,root,0600 -w
Setup Hadoop
For each of the created VMs, follow these steps:
Log as root and set the following variables:
HADOOP_PREFIX=/usr/local/hadoop-2.7.2
HADOOP_HOME=$HADOOP_PREFIX
HADOOP_COMMON_HOME=$HADOOP_PREFIX
HADOOP_CONF_DIR=$HADOOP_PREFIX/etc/hadoop
HADOOP_HDFS_HOME=$HADOOP_PREFIX
HADOOP_MAPRED_HOME=$HADOOP_PREFIX
HADOOP_YARN_HOME=$HADOOP_PREFIX
23 https://slider.incubator.apache.org
Public
Page 29 of 40
OpenMinTeD Platform services distribution specification
•••
$HADOOP_CONF_DIR/yarn-site.xml
<configuration>
<property>
<name>yarn.nodemanager.container-executor.class</name>
<value>org.apache.hadoop.yarn.server.nodemanager.DockerContainerExecutor</value>
</property>
<property>
<name>yarn.nodemanager.docker-container-executor.exec-name</name>
<value>docker</value>
</property>
<!-- For the cluster -->
<property>
<name>yarn.resourcemanager.hostname</name>
<value>resourcemanager.<hadoop 1></value>
<description>The hostname of the RM.</description>
</property>
</configuration>
$HADOOP_CONF_DIR/hdfs-site.xml
<configuration>
<property>
<name>dfs.datanode.data.dir</name>
<value>file:///usr/local/hadoop-2.7.2/hdfs/datanode</value>
<description>Comma separated list of paths on the local filesystem of a DataNode where it should store its
blocks.</description>
</property>
<property>
<name>dfs.namenode.name.dir</name>
<value>file:///usr/local/hadoop-2.7.2/hdfs/namenode</value>
<description>Path on the local filesystem where the NameNode stores the namespace and transaction logs
persistently.</description>
Public
Page 30 of 40
OpenMinTeD Platform services distribution specification
•••
</property>
</configuration>
$HADOOP_CONF_DIR/core-site.xml
<configuration>
<property>
<name>fs.defaultFS</name>
<value>hdfs://namenode.<hadoop 1 IP></value>
<description>NameNode URI</description>
</property>
</configuration>
Start Hadoop nodes and managers
#Start HDFS daemons
# Format the namenode directory (DO THIS ONLY ONCE, THE FIRST TIME)
# ONLY ON THE NAMENODE NODE
$HADOOP_PREFIX/bin/hdfs namenode -format
# Start the namenode daemon
# ONLY ON THE NAMENODE NODE
$HADOOP_PREFIX/sbin/hadoop-daemon.sh start namenode
# Start the datanode daemon
# ON ALL SLAVES
$HADOOP_PREFIX/sbin/hadoop-daemon.sh start datanode
## Start YARN daemons
# Start the resourcemanager daemon
# ONLY ON THE RESOURCEMANAGER NODE
$HADOOP_PREFIX/sbin/yarn-daemon.sh start resourcemanager
Public
Page 31 of 40
OpenMinTeD Platform services distribution specification
•••
# Start the nodemanager daemon
# ON ALL SLAVES
$HADOOP_PREFIX/sbin/yarn-daemon.sh start nodemanager
Test it
# $HADOOP_PREFIX/bin/hadoop jar \
$HADOOP_PREFIX/share/hadoop/mapreduce/hadoop-mapreduce-examples-2.7.2.jar teragen \
-Dmapreduce.map.env="yarn.nodemanager.docker-container-executor.image-name=sequenceiq/hadoopdocker:2.4.1" \
-Dyarn.app.mapreduce.am.env="yarn.nodemanager.docker-container-executor.imagename=sequenceiq/hadoop-docker:2.4.1" \
1000 teragen_out_dir
# $HADOOP_PREFIX/bin/hdfs dfs -cat teragen_out_dir/*
Install slider
Unfortunately, you will need Apache Maven and you then need to download Slider and install it with
maven.
# apt install maven
# export MAVEN_OPTS="-Xms256m -Xmx512m -Djava.awt.headless=true"
# git clone https://git-wip-us.apache.org/repos/asf/incubator-slider.git -b develop
# cd incubator-slider
# mvn clean install
# mvn clean site:site site:stage package -DskipTests
# export SLIDER_HOME=/root/incubator-slider/slider-assembly/target/slider-0.92.0-incubating-SNAPSHOTall/slider-0.92.0-incubating-SNAPSHOT
Run with slider
Slider builds app packages based on three config files: appConfig.json (application specific overrides),
metainfo.json (more generic build configs) and resources.json (deployment instructions to yarn).
metainfo.json
{
"schemaVersion": "2.1",
"application": {
Public
Page 32 of 40
OpenMinTeD Platform services distribution specification
•••
"name": "WORDPRESS",
"components": [
{
"name": "WORDPRESS",
"type": "docker",
"dockerContainers": [
{
"name": "wordpress",
"commandPath": "/usr/bin/docker",
"image": "wordpress",
"ports": [{
"containerPort" : "8080"
}]
}]
},
{
"name": "MYSQL",
"type": "docker",
"dockerContainers": [
{
"name": "mysql",
"commandPath": "/usr/bin/docker",
"image": "mysql:5.5",
"ports": [{
"containerPort" : "3306",
"hostPort": "3306"
}]
}]
}]
}
}
resources.json:
{
Public
Page 33 of 40
OpenMinTeD Platform services distribution specification
•••
"metadata": { },
"global": { },
"components": {
"slider-appmaster": { },
"MYSQL": {
"yarn.role.priority": "1",
"yarn.component.instances": "1",
"yarn.memory": "512"
},
"WORDPRESS": {
"yarn.role.priority": "2",
"yarn.component.instances": "1",
"yarn.memory": "512"
}
}
}
Now, let's deploy the app:
# $SLIDER_HOME/bin/slider create SliPress --template appConfig.json --metainfo metainfo.json --resources
resources.json
Public
Page 34 of 40
OpenMinTeD Platform services distribution specification
•••
7. Annex 3: Run Hadoop YARN with GRNET ~orka
Introduction
The following guides document how ~okeanos users can access and use ~orka services. Focus is given
on the available images, the creation of a cluster from the GUI and the respective CLI commands. Full
documentation about ~orka is provided from https://okeanos.grnet.gr/platforms/orka/ and
https://escience.grnet.gr/static/mkdocs/site/introduction/index.html.
Information about ~orka images
In order to view all the available Hadoop distributions (images) the user selects, «Hadoop Images
Components», displays in a tabular format all the components comprising the pre-cooked Hadoop
images, their respective versions and their web homepages for further information:
Figure 2: Component of a selected Hadoop image
~orka overview page
After successful authentication, the user is redirected to the landing page, with an overview of all the
infrastructures created inside the application. There is information about Hadoop (YARN) clusters,
Virtual Research Environments, and Reproducible Experiments in a grid format:
Public
Page 35 of 40
OpenMinTeD Platform services distribution specification
•••
Figure 3: List of all created Hadoop clusters
Creating a Hadoop cluster
The basic function of ~orka is to create a multi-node Hadoop cluster, through a friendly UI:
Public
Page 36 of 40
OpenMinTeD Platform services distribution specification
•••
Figure 4: Snapshot of ~orka Hadoop creation form
The following is a screenshot from a HUE (Hue User Interface) dashboard:
Public
Page 37 of 40
OpenMinTeD Platform services distribution specification
•••
Figure 5: List of all jobs in Hadoop cluster
~orka CLI examples
The following CLI command illustrate how to create a 10-node Hadoop cluster (each node with 4CPU
4GM RAM and 40GB of disk space):
$ orka create OMTD_Hadoop 10 4 4096 40 4 4096 40 Standard <project_name> --image=Ecosystem-on-Hue3.8.0
The following command lists info about this cluster (the create command has returned an id):
$ orka info <cluster_id>
To scale the cluster, use this command:
$ orka node add <cluster_id>
Command to start the Hadoop cluster:
$ orka hadoop start <cluster_id>
C
ommand to stop the Hadoop cluster:
$ orka hadoop stop <cluster_id>
Command for putting a local file to HDFS:
$ orka file put <cluster_id> <remote_http_or_ftp_url> <hdfs_file_path>
Public
Page 38 of 40
OpenMinTeD Platform services distribution specification
•••
Command for putting a Pithos+ file to HDFS:
$ orka file put <cluster_id> pithos://<pithos_file_path> <hdfs_file_path>
Public
Page 39 of 40
OpenMinTeD Platform services distribution specification
•••
8. References
1. D4.3 – OpenMinTeD Functional Specifications
2. D8.1 - Infrastructure Operation Report
Public
Page 40 of 40
Was this manual useful for you? yes no
Thank you for your participation!

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

Download PDF

advertisement