Overview

The execjob_postsuspend hook event will be executed after a pbs_mom successfully suspends a job. The execjob_preresume hook event will be executed before a pbs_mom attempts to resume a job. This will allow for notifying changes in resource allocation to external consumers (e.g. cgroups).

Discussion here: http://community.pbspro.org/t/two-new-hook-events-execjob-postsuspend-and-execjob-preresume/1469

Technical Details

Interface 1: new mom hook event: execjob_postsuspend

• Python constant: pbs.EXECJOB_POSTSUSPEND
• Event Parameters: 
  • pbs.event().job - This is a pbs.job object representing the job that has been suspended. This job object cannot be modified under this hook.
  • pbs.event().vnode_list[] - This is a dictionary of pbs.vnode objects, keyed by vnode name, listing the vnodes that are assigned to the job. The vnode objects in the vnode_list cannot be modified.
• Hook Attributes:
  • fail_action: This hook will not allow a fail action to be set.
  • user: This hook will only allow the value "pbsadmin".
• Details:
  • An execjob_postsuspend hook is executed by a mom after the job has been successfully suspended on her node.

  • A call to pbs.event().accept() means the hook code has executed cleanly.

  • A call to pbs.event().reject() means the hook code was not able to fully accomplish its task

    • Note: this will not cause the job to be resumed.
    • Keeping with hook design, if one execjob_postsuspend hook is rejected, the other execjob_postsuspend hooks with a higher order value will not run.

  • If the execjob_postsuspend hook script encounters an unexpected error causing an unhandled exception, or times out due to the hook's alarm setting, the hook will act similar to a pbs.event().reject().

• Internal Details
  • The primary mom does not need to know the outcome of the sister mom's hook events. Therefore, the primary mom does not need to request the sisters to run the event.
  • Instead, each mom should run this event as part of their suspending.
• Additional details:
  
  • Other than adding a step to run the hooks, this hook event does not change the flow of the Suspend/Resume.
• Consumer:
  • Hooks like cgroups that need to take action when resource allocation changes.
    • When a job is suspended, the cgroup needs to be removed/modified. Otherwise, when the preempting job tries to run, the cgroup will be rejected.

Interface 2: new mom hook event: execjob_preresume

• Python constant: pbs.EXECJOB_PRERESUME
• Event Parameters: 
  • pbs.event().job - This is a pbs.job object representing the job that will be resumed. This job object cannot be modified under this hook.
  • pbs.event().vnode_list[] - This is a dictionary of pbs.vnode objects, keyed by vnode name, listing the vnodes that are assigned to the job. The vnode objects in the vnode_list cannot be modified.
• Hook Attributes:
  • fail_action: This hook will not allow a fail action to be set.
  • user: This hook will only allow the value "pbsadmin".
• Details:
  • An execjob_preresume hook is executed by the primary mom when a request to resume the job is received.

  • An execjob_preresume hook is executed by the sister mom when a request from the primary mom to resume the job's tasks is received.

  • A call to pbs.event().accept() means the hook code has executed cleanly.

  • A call to pbs.event().reject() means the hook code was not able to fully accomplish its task

    • Note: this will prevent all MoMs from resuming jobs.
    • Keeping with hook design, if one execjob_preresume hook is rejected, the other execjob_preresume hooks with a higher order value will not run.

  • If the execjob_preresume hook script encounters an unexpected error causing an unhandled exception, or times out due to the hook's alarm setting, the hook will act similar to a pbs.event().reject().

    • Note: this will prevent all MoMs from resuming jobs.
• Internal Design:
  
  • The MS will complete the event first. If it is not rejected, the sisters will then run their hooks.
  • All moms must accept the event before the job can be resumed.
• Consumer:
  • Hooks like cgroups that need to take action when resource allocation changes.
    • Because the cgroups are cleaned up on suspension, it has to be recreated/modified when the job is resumed. Otherwise, the job will not have the resources it requested.
• Caveats
  • Current behavior shows that when a PBS_BATCH_SignalJob to resume a job is rejected by the mom, the server starts another scheduling cycle. If the scheduler says it can still be resumed, it will try again. If the execjob_preresume hook always rejects, there is nothing preventing this loop. Again, this is consistent with current behavior, but now it's easier to enter this loop.