The gLite Workload Management System Alessandro Maraschini alessandro'maraschinidatamat'it

1
The gLite Workload Management SystemAlessandro
Maraschini alessandro.maraschini_at_datamat.it
OGF06, Manchester, UK 2007, May 9
2
Contents

• JRA1 WMS
• JRA1 overview, scope
• WMS, institutes, task, components
• JDL
• Language overview
• JobTypes single, compounds, workflows
• News
• New Functionalities
• Latest Activities
• Tests
• Middleware Testing Activities Results
• Future Plans conclusions
• Workflow Activities
• Ongoing Implementations Activities
• Future Implementations Activities
• WMS and WfMS

3
Introduction gLite WMS

• Workload Management System (WMS)
• Part of Joint and Research Activity 1 (JRA1)
• Institutes involved
• INFN
• Datamat
• CESNET
• Provides Distribution and management of tasks
  across resources available on a Grid
• Accept a request of execution of a Job from a
  client
• Find appropriate resources to satisfy the Job
• Follow the Job until completion.
• Different aspects of job management are
  accomplished by different WMS components
• Implemented as different processes
• Communicating via data structures stored on disk
  avoid data losses
• Job associated to user credentials- operations
  done on behalf of the user

4
WMS Architecture JRA1 core components

• WMProxy
• Accept Request from User
• Check Authentication/Authorization
• Set up Local File System
• provide access to the WMS
• Forward request to WM
• Workload Manager (WM)
• Accept and satisfy requests for job management
  coming from its clients
• Forward request to appropriate Computing Element
  (CE) for execution
• Logging Bookkeeping (LB)
• Tracks jobs in terms of events gathered from
  various gLite components
• The server processes the incoming events to give
  a higher level view on the job states (e.g.
  Submitted, Running, Done)

5
WMS Architecture overview
6
JDL overview

• Job Description Language (JDL)
• gLite approach to Request Description
• classads-based language
• Fully extensible flexible high-level language
• Allow the user to provide job execution needed
  information
• Characteristics of the application
  (Executable,Arguments,Input/Output Sandbox
  files,...)
• Requirements/preferences about resources (
  Computational, storage)
• Customized hints for gLite WMS on how to handle
  the application (number of retries, proxy
  renewal, ...)
• Supported Job Types
• Single Jobs
• Compound Jobs
• Workflows (DAGS)
• Collections, Parametrics

7
JDL Single Types

• Single Jobs
• Normal single and simple batch job with no
  peculiar requirements
• MPICH a parallel application to be run on the
  nodes of a cluster using the MPICH implementation
  of the message passing interface ( new flavour of
  MPI modifiche per estendere supportare IN FUTURE
  )
• Interactive a job whose standard streams are
  forwarded to the submitting client, which can
  actually interact and steer the job execution by
  providing real-time input information
• Previously Supported Jobs
• Deprecation due Lack of feedback
• Not anymore supported
• Checkpointable Jobs
• Partitionable Jobs

8
JDL Compound Jobs

• Definition
• Aggregate of Normal Jobs
• Benefits
• One Shot submission for (up to thousands of) jobs
• Submission time reduction
• Single call to WMProxy server
• Single AuthN and AuthZ process
• Sharing of files between jobs
• Single Identification to manage all jobs (father
  Job)

9
JDL Compound Types

• Compound Jobs Workflows
• Implemented as Directed Acyclic Graphs (DAGS)
• Set of jobs where the input, output or execution
  of one of more jobs may depend on one or more
  other jobs
• Dependencies represent time constraints a child
  cannot start before all parents have successfully
  completed

10
JDL Compound Types

• Compound Jobs Parametrics
• Parameterized description of a Job
• automatically converted on WMS side
• generates a (possibly) huge number of (similar)
  jobs

11
JDL Compound Types

• Compound Jobs Collections
• A set of possibly eterogeneus jobs that can be
  specified within a single JDL description
• No dependencies among the specified jobs
• Jobs executed independently among the grid

12
New Functionalities WMProxy

• WMProxy server
• Replaced the old C based socket connection
  sevice
• implements an interoperable interface
• Web Service based
• SOA conformance
• WS-I compliance
• provided new operations
• WMProxy client
• Provided C based WMS command-line User
  Interface (UI), which executes all the needed
  operation automatically
• Provided multi language (C, Java and Python)
  provided APIs
• Fell free to implement your own client with your
  desired language

13
New Functionalities ICE- CREAM

• WMS Job submission supported resources
• LCG Computing Element
• gLite condor-based Computing Element
• Moreover, the recent introduction of ICE
• intermediate layer service
• allows the WMS to directly send operations to
  CREAM
• Computing Resource Execution And Management
  Service
• asynchronously receive notifications about job
  status changes

14
New Functionalities Sandbox Files

• Sandbox Archiving and Sharing
• Job sandbox files can be automatically compressed
• Different jobs can share the same sandbox,
• dramatically reduced network traffic
• allowed the user to save time and bandwidth
• Sandbox Remote Specification
• User can store files directly on a remote machine
• No intermediate copies workernode will download
  directly
• Reduced server load
• Supported File Transfer
• Full support (submssion output file retrieval)
  for protocols
• gridftp
• https

15
New Functionalities Bulk-MM

• Bulk-Matchmaking
• Allow single Matchmaking of similar jobs in one
  shot
• Jobs equivalence based upon submitting
  significant attributes
• Target Jobs Bunch of Independent Jobs
• Mainly Collections and Parametrics
• Originally managed with DAGMan
• Saved time resources
• Improved System stability and performances
• non-final status jobs decreased from 5 to 0,3
  (see next slides)

16
Other New Functionalities

• Service Discovery
• provide additional information by performing
  queries to external databases of different kinds
  (RGMA, BDII)
• Client side
• Queries for available WMProxy Endpoints on the
  net
• Do not need user commands manual reconfiguration
• Server side
• Queries for available LB servers where to Log Job
  information
• Job Files Perusal
• Perform a monitoring activity on the actual
  output files produced by a job during its
  lifecycle
• Add important pieces of information not available
  by simple status monitoring and that were before
  available only at job completion

17
New Activities

• New platforms widely deployed on the
  infrastructure
• In particular Scientific Linux 4 and 64-bit
  architectures
• Migration to ETICS build system
• More flexible, in particular in
• Addresses multiple platform support
• almost impossible using the old gLite build
  system
• All WMS components build achieved
• Client side manual installation fully working
• Ongoing activity Integration
• Software not yet fully deployed
• Server side installation not yet available (will
  be in short term..)

18
Test Result

• Intense testing and constant bug fixing
  activities have been performed over the last
  months
• Improved job submission rate
• Improved service stability
• New Functionalities tested and adopted
• Production quality test Results
• 16K jobs/day over one week of submissions
• No manual intervention on server
• Stable memory usage
• 0.3 of jobs in non-final state
• Aborted jobs mostly due to expired user
  credentials

19
Test Result
20
gLite WMS Ongoing Restructuring

• gLite Restructuring
• All activities stopped for 6 months
• improving usability portability
• Multi platform (Structural changes needed)
• Cleaning up sections that cause build and porting
  difficulties
• Removing/Reducing Dependencies on external
  software
• Objectives
• Easier Service maintainance and Usage
• Will increase stability and throughput
• Toward a gLighter User Interface
• Identify and remove all unnecessary dependencies

21
gLite WMS Future

• Improving Logging and Error Reporting
• Windows working prototype
• gLite porting on MS
• Improving interoperability
• Supercomputing 06 working prototype
• Basic Execution Service (BES)
• Job Submission Description Language (JSDL)

22
WfMS and gLite WMS

• Possible integration with external existing
  Workflow managers
• Triana, GWES, Taverna, etc
• Still to be discussed and planned for EGEE III
• Moreover, Workflow Mangement System (WfMS)
  Architecture Proposal for WMS
• Running on top of gLite Middleware
• Grid Middleware Undependent
• Abstract and Generic Representation
• Translation mechanisms from different language
  front ends
• Will be exposed/discussed at next CoreGrid forum