Knowledge Center         Contents    Previous  Next    Index  
Platform Computing Corp.

Chunk Job Dispatch

Contents

About Job Chunking

LSF supports job chunking, where jobs with similar resource requirements submitted by the same user are grouped together for dispatch. The CHUNK_JOB_SIZE parameter in lsb.queues and lsb.applications specifies the maximum number of jobs allowed to be dispatched together in a chunk job.

Job chunking can have the following advantages:

All of the jobs in the chunk are dispatched as a unit rather than individually. Job execution is sequential, but each chunk job member is not necessarily executed in the order it was submitted.

restriction:  
You cannot auto-migrate a suspended chunk job member.

Chunk job candidates

Jobs with the following characteristics are typical candidates for job chunking:

Running jobs with these characteristics without chunking can underutilize resources because LSF spends more time scheduling and dispatching the jobs than actually running them.

Configuring a special high-priority queue for short jobs is not desirable because users may be tempted to send all of their jobs to this queue, knowing that it has high priority.

Configure Chunk Job Dispatch

CHUNK_JOB_SIZE (lsb.queues)

By default, CHUNK_JOB_SIZE is not enabled.

  1. To configure a queue to dispatch chunk jobs, specify the CHUNK_JOB_SIZE parameter in the queue definition in lsb.queues.
  2. For example, the following configures a queue named chunk, which dispatches up to 4 jobs in a chunk:

    Begin Queue
    QUEUE_NAME     = chunk
    PRIORITY       = 50
    CHUNK_JOB_SIZE = 4
    End Queue 
    

Postrequisites: After adding CHUNK_JOB_SIZE to lsb.queues, use badmin reconfig to reconfigure your cluster.

Chunk jobs and job throughput

Throughput can deteriorate if the chunk job size is too big. Performance may decrease on queues with CHUNK_JOB_SIZE greater than 30. You should evaluate the chunk job size on your own systems for best performance.

CHUNK_JOB_SIZE (lsb.applications)

By default, CHUNK_JOB_SIZE is not enabled. Enabling application-level job chunking overrides queue-level job chunking.

  1. To configure an application profile to chunk jobs together, specify the CHUNK_JOB_SIZE parameter in the application profile definition in lsb.applications.
  2. Specify CHUNK_JOB_SIZE=1 to disable job chunking for the application. This value overrides chunk job dispatch configured in the queue.

Postrequisites: After adding CHUNK_JOB_SIZE to lsb.applications, use badmin reconfig to reconfigure your cluster.

CHUNK_JOB_DURATION (lsb.params)

If CHUNK_JOB_DURATION is defined in the file lsb.params, a job submitted to a chunk job queue is chunked under the following conditions:

and the values of the CPU limit, run time limit, and run time estimate are all less than or equal to the CHUNK_JOB_DURATION.

Jobs are not chunked if:

The value of CHUNK_JOB_DURATION is displayed by bparams -l.

  1. After adding CHUNK_JOB_DURATION to lsb.params, use badmin reconfig to reconfigure your cluster.
  2. By default, CHUNK_JOB_DURATION is not enabled.

Restrictions on chunk jobs

CHUNK_JOB_SIZE is ignored and jobs are not chunked under the following conditions:

Jobs submitted with the following bsub options are not chunked; they are dispatched individually:

Submitting and Controlling Chunk Jobs

When a job is submitted to a queue or application profile configured with the CHUNK_JOB_SIZE parameter, LSF attempts to place the job in an existing chunk. A job is added to an existing chunk if it has the same characteristics as the first job in the chunk:

If a suitable host is found to run the job, but there is no chunk available with the same characteristics, LSF creates a new chunk.

Resources reserved for any member of the chunk are reserved at the time the chunk is dispatched and held until the whole chunk finishes running. Other jobs requiring the same resources are not dispatched until the chunk job is done.

For example, if all jobs in the chunk require a software license, the license is checked out and each chunk job member uses it in turn. The license is not released until the last chunk job member is finished running.

WAIT status

When sbatchd receives a chunk job, it does not start all member jobs at once. A chunk job occupies a single job slot. Even if other slots are available, the chunk job members must run one at a time in the job slot they occupy. The remaining jobs in the chunk that are waiting to run are displayed as WAIT by bjobs. Any jobs in WAIT status are included in the count of pending jobs by bqueues and busers. The bhosts command shows the single job slot occupied by the entire chunk job in the number of jobs shown in the NJOBS column.

The bhist -l command shows jobs in WAIT status as Waiting ...

The bjobs -l command does not display a WAIT reason in the list of pending jobs.

Controlling chunk jobs

Job controls affect the state of the members of a chunk job. You can perform the following actions on jobs in a chunk job:

Action (Command)
Job State
Effect on Job (State)
Suspend (bstop)
PEND
Removed from chunk (PSUSP)
 
RUN
All jobs in the chunk are suspended
(NRUN -1, NSUSP +1)
 
USUSP
No change
 
WAIT
Removed from chunk (PSUSP)
Kill (bkill)
PEND
Removed from chunk (NJOBS -1, PEND -1)
 
RUN
Job finishes, next job in the chunk starts if one exists (NJOBS -1, PEND -1)
 
USUSP
Job finishes, next job in the chunk starts if one exists (NJOBS -1, PEND -1, SUSP -1, RUN +1)
 
WAIT
Job finishes (NJOBS-1, PEND -1)
Resume (bresume)
USUSP
Entire chunk is resumed (RUN +1, USUSP -1)
Migrate (bmig)
WAIT
Removed from chunk
Switch queue (bswitch)
RUN
Job is removed from the chunk and switched; all other WAIT jobs are requeued to PEND
 
WAIT
Only the WAIT job is removed from the chunk and switched, and requeued to PEND
Checkpoint (bchkpnt)
RUN
Job is checkpointed normally
Modify (bmod)
PEND
Removed from the chunk to be scheduled later

Migrating jobs with bmig changes the dispatch sequence of the chunk job members. They are not redispatched in the order they were originally submitted.

Rerunnable chunk jobs

If the execution host becomes unavailable, rerunnable chunk job members are removed from the queue and dispatched to a different execution host.

See Chapter 30, "Job Requeue and Job Rerun" for more information about rerunnable jobs.

Checkpointing chunk jobs

Only running chunk jobs can be checkpointed. If bchkpnt -k is used, the job is also killed after the checkpoint file has been created. If chunk job in WAIT state is checkpointed, mbatchd rejects the checkpoint request.

See Chapter 31, "Job Checkpoint, Restart, and Migration" for more information about checkpointing jobs.

Fairshare policies and chunk jobs

Fairshare queues can use job chunking. Jobs are accumulated in the chunk job so that priority is assigned to jobs correctly according to the fairshare policy that applies to each user. Jobs belonging to other users are dispatched in other chunks.

TERMINATE_WHEN job control action

If the TERMINATE_WHEN job control action is applied to a chunk job, sbatchd kills the chunk job element that is running and puts the rest of the waiting elements into pending state to be rescheduled later.

Enforce resource usage limits on chunk jobs

By default, resource usage limits are not enforced for chunk jobs because chunk jobs are typically too short to allow LSF to collect resource usage.

  1. To enforce resource limits for chunk jobs, define LSB_CHUNK_RUSAGE=Y in lsf.conf. Limits may not be enforced for chunk jobs that take less than a minute to run.

Platform Computing Inc.
www.platform.com
Knowledge Center         Contents    Previous  Next    Index