doc/html/checkpoint_blcr.shtml - SchedMD/slurm - Git at Google

 <!--#include virtual="header.txt"-->

 <h1><a name="top">SLURM Checkpoint/Restart with BLCR</a></h1>

 <h2>Overview</h2>
 <p>SLURM version 2.0 has been integrated with
 <a href="https://ftg.lbl.gov/CheckpointRestart/CheckpointRestart.shtml">
 Berkeley Lab Checkpoint/Restart (BLCR)</a> in order to provide automatic
 job checkpoint/restart support.
 Functionality provided includes:
 <ol>
 <li>Checkpoint of whole batch jobs in addition to job steps</li>
 <li>Periodic checkpoint of batch jobs and job steps</li>
 <li>Restart execution of batch jobs and job steps from checkpoint files</li>
 <li>Automatically  requeue and restart the execution of batch jobs upon
 node failure</li>
 </ol></p>

 <h2>User Commands</h2>

 <p>The following documents SLURM changes specific to BLCR support.
 Baic familiarity with SLURM commands is assumed.</p>

 <h3>srun</h3>

 <p>Several options have been added to support checkpoint restart:</p>
 <ul>
 <li><b>--checkpoint</b>: Specify the interval between periodic checkpoint
 of a job step, in seconds</li>
 <li><b>--checkpoint-dir</b>:Specify the directory when the checkpoint image
 files of a job step will be stored.
 The default value is the current working directory.
 Checkpoint files will be of the form <i>"&lt;job_id&gt;.ckpt"</i> for jobs
 and <i>"&lt;job_id&gt;.&lt;step_id&gt;.ckpt"</i> for job steps.</li>
 <li><b>--restart-dir</b>: Specify the directory when the checkpoint image
 files of a job step will be read from</li>
 </li>
 </ul>

 <p>Environment variables are available for all of these options:</p>
 <ul>
 <li<b>SLURM_CHECKPOINT</b> is equivalent to <b>--checkpoint</b>:</li>
 <li><b>SLURM_CHECKPOINT_DIR</b> is equivalent to <b>--checkpoint-dir</b></li>
 <li><b>SLURM_RESTART_DIR</b> is equivalent to <b>--restart-dir</b></li>
 </li>
 </ul>
 <p>The environment variable <b>SLURM_SRUN_CR_SOCKET</b> is used for job step
 logic to interact with the <b>srun_cr</b> command.</p>

 <h3>srun_cr</h3>

 <p>This is a wrapper program for use with SLURM's <b>checkpoint/blcr</b>
 plugin to checkpoint/restart tasks launched by srun.
 The design of <b>srun_cr</b> is inspired by <b>mpiexec_cr</b> from MVAPICH2 and
 <b>cr_restart</b> form BLCR.
 It is a wrapper around the <b>srun</b> command to enable batch job
 checkpoint/restart support when used with SLURM's <b>checkpoint/blcr</b> plugin.

 <p>The <b>srun_cr</b> execute line options are identical to those of the
 <b>srun</b> command.
 See "man srun" for details.</p>

 <p>After initialization, <b>srun_cr</b> registers a thread context callback
 function.
 Then it forks a process and executes "cr_run --omit srun" with its arguments.
 <b>cr_run</b> is employed to exclude the <b>srun</b> process from being dumped
 upon checkpoint.
 All catchable signals except SIGCHLD sent to <b>srun_cr</b> will be forwarded
 to the child <b>srun</b> process.
 SIGCHLD will be captured to mimic the exit status of <b>srun</b> when it exits.
 Then <b>srun_cr</b> loops waiting for termination of tasks being launched
 from <b>srun</b>.</p>

 <p>The step launch logic of SLURM is augmented to check if <b>srun</b> is
 running under <b>srun_cr</b>.
 If true, the environment variable <b>SURN_SRUN_CR_SOCKET</b> should be present,
 the value of which is the address of a Unix domain socket created and listened
 to be <b>srun_cr</b>.
 After launching the tasks, <b>srun</b> tires to connect to the socket and sends
 the job ID, step ID and the nodes allocated to the step to <b>srun_cr</b>.</p>

 <p>Upon checkpoint, </b>srun_cr</b> checks to see if the tasks have been launched.
 If not </b>srun_cr</b> first forwards the checkpoint request to the tasks by
 calling the SLURM API <b>slurm_checkpoint_tasks()</b> before dumping its process
 context.</p>

 <p>Upon restart, <b>srun_cr</b> checks to see if the tasks have been previously
 launched and checkpointed.
 If true, the environment variable </b>SLURM_RESTART_DIR</b> is set to the
 directory of the checkpoint image files of the tasks.
 Then <b>srun</b> is forked and executed again.
 The environment variable will be used by the <b>srun</b> command to restart
 execution of the tasks from the previous checkpoint.</p>

 <h3>sbatch</h3>

 <p>Several options have been added to support checkpoint restart:</p>
 <ul>
 <li><b>--checkpoint</b>: Specify the interval between periodic checkpoint
 of a batch job, in seconds</li>
 <li><b>--checkpoint-dir</b>:Specify the directory when the checkpoint image
 files of a batch job will be stored.
 The default value is the current working directory.
 Checkpoint files will be of the form <i>"&lt;job_id&gt;.ckpt"</i> for jobs
 and <i>"&lt;job_id&gt;.&lt;step_id&gt;.ckpt"</i> for job steps.</li>
 </li>
 </ul>

 <p>Environment variables are available for all of these options:</p>
 <ul>
 <li<b>SLURM_CHECKPOINT</b> is equivalent to <b>--checkpoint</b>:</li>
 <li><b>SLURM_CHECKPOINT_DIR</b> is equivalent to <b>--checkpoint-dir</b></li>
 </li>
 </ul>

 <h3>scontrol</h3>

 <p><b>scontrol</b> is used to initiate checkpoint/restart requests.</p>
 <ul>
 <li><b>scontrol checkpoint create <i>jobid</i> [ImageDir=<i>dir</i>]
 [MaxWait=<i>seconds</i>]</b><br>
 Requests a checkpoint on a specific job.
 For backward compatibility, if a job id is specified, all job steps of
 it are checkpointed.
 If a batch job id is specified, the entire job is checkpointed including
 the batch shell and all running tasks of all job steps.
 Upon checkpoint, the task launch command must forward the requests to
 tasks it launched.
 <ul>
 <li><b>ImageDir</b> specifies the directory in which to save the checkpoint
 image files. If specified, this takes precedence over any <b>--checkpoint-dir</b>
 option specified when the job or job step were submitted.</li>
 <li><b>MaxWait</b> specifies the maximum time permitted for a checkpoint
 request to complete. The request will be considered failed if not
 completed in this time period.</li>
 </li>
 </ul>

 <li><b>scontrol checkpoint create <i>jobid.stepid</i> [ImageDir=<i>dir</i>]
 [MaxWait=<i>seconds</i>]</b><br>
 Requests a checkpoint on a specific job step.</li>

 <li><b>scontrol checkpoint restart <i>jobid</i> [ImageDir=<i>dir</i>]
 [StickToNodes]</b><br>
 Restart a previously checkpointed batch job.
 <ul>
 <li><b>ImageDir</b> specifies the directory from which to read the checkpoint
 image files.</li>
 <li><b>StickToNodes</b> specifies that the job should be restarted on the
 same set of nodes from which it was previously checkpointed.</li>
 </ul></li>
 </ul>

 <h2>Configuration</h2>

 <p>The following SLURM configuration parameter has been added:</p>
 <ul>
 <li><b>JobCheckpointDir</b> specified the default directory for storing
 or reading job checkpoint files</li>
 </ul>

 <p class="footer"><a href="#top">top</a></p>

 <p style="text-align:center;">Last modified 11 March 2009</p>

 <!--#include virtual="footer.txt"-->
	<!--#include virtual="header.txt"-->

	<h1><a name="top">SLURM Checkpoint/Restart with BLCR</a></h1>

	<h2>Overview</h2>
	<p>SLURM version 2.0 has been integrated with
	<a href="https://ftg.lbl.gov/CheckpointRestart/CheckpointRestart.shtml">
	Berkeley Lab Checkpoint/Restart (BLCR)</a> in order to provide automatic
	job checkpoint/restart support.
	Functionality provided includes:
	<ol>
	<li>Checkpoint of whole batch jobs in addition to job steps</li>
	<li>Periodic checkpoint of batch jobs and job steps</li>
	<li>Restart execution of batch jobs and job steps from checkpoint files</li>
	<li>Automatically requeue and restart the execution of batch jobs upon
	node failure</li>
	</ol></p>

	<h2>User Commands</h2>

	<p>The following documents SLURM changes specific to BLCR support.
	Baic familiarity with SLURM commands is assumed.</p>

	<h3>srun</h3>

	<p>Several options have been added to support checkpoint restart:</p>
	<ul>
	<li><b>--checkpoint</b>: Specify the interval between periodic checkpoint
	of a job step, in seconds</li>
	<li><b>--checkpoint-dir</b>:Specify the directory when the checkpoint image
	files of a job step will be stored.
	The default value is the current working directory.
	Checkpoint files will be of the form <i>"<job_id>.ckpt"</i> for jobs
	and <i>"<job_id>.<step_id>.ckpt"</i> for job steps.</li>
	<li><b>--restart-dir</b>: Specify the directory when the checkpoint image
	files of a job step will be read from</li>
	</li>
	</ul>

	<p>Environment variables are available for all of these options:</p>
	<ul>
	<li<b>SLURM_CHECKPOINT</b> is equivalent to <b>--checkpoint</b>:</li>
	<li><b>SLURM_CHECKPOINT_DIR</b> is equivalent to <b>--checkpoint-dir</b></li>
	<li><b>SLURM_RESTART_DIR</b> is equivalent to <b>--restart-dir</b></li>
	</li>
	</ul>
	<p>The environment variable <b>SLURM_SRUN_CR_SOCKET</b> is used for job step
	logic to interact with the <b>srun_cr</b> command.</p>

	<h3>srun_cr</h3>

	<p>This is a wrapper program for use with SLURM's <b>checkpoint/blcr</b>
	plugin to checkpoint/restart tasks launched by srun.
	The design of <b>srun_cr</b> is inspired by <b>mpiexec_cr</b> from MVAPICH2 and
	<b>cr_restart</b> form BLCR.
	It is a wrapper around the <b>srun</b> command to enable batch job
	checkpoint/restart support when used with SLURM's <b>checkpoint/blcr</b> plugin.

	<p>The <b>srun_cr</b> execute line options are identical to those of the
	<b>srun</b> command.
	See "man srun" for details.</p>

	<p>After initialization, <b>srun_cr</b> registers a thread context callback
	function.
	Then it forks a process and executes "cr_run --omit srun" with its arguments.
	<b>cr_run</b> is employed to exclude the <b>srun</b> process from being dumped
	upon checkpoint.
	All catchable signals except SIGCHLD sent to <b>srun_cr</b> will be forwarded
	to the child <b>srun</b> process.
	SIGCHLD will be captured to mimic the exit status of <b>srun</b> when it exits.
	Then <b>srun_cr</b> loops waiting for termination of tasks being launched
	from <b>srun</b>.</p>

	<p>The step launch logic of SLURM is augmented to check if <b>srun</b> is
	running under <b>srun_cr</b>.
	If true, the environment variable <b>SURN_SRUN_CR_SOCKET</b> should be present,
	the value of which is the address of a Unix domain socket created and listened
	to be <b>srun_cr</b>.
	After launching the tasks, <b>srun</b> tires to connect to the socket and sends
	the job ID, step ID and the nodes allocated to the step to <b>srun_cr</b>.</p>

	<p>Upon checkpoint, </b>srun_cr</b> checks to see if the tasks have been launched.
	If not </b>srun_cr</b> first forwards the checkpoint request to the tasks by
	calling the SLURM API <b>slurm_checkpoint_tasks()</b> before dumping its process
	context.</p>

	<p>Upon restart, <b>srun_cr</b> checks to see if the tasks have been previously
	launched and checkpointed.
	If true, the environment variable </b>SLURM_RESTART_DIR</b> is set to the
	directory of the checkpoint image files of the tasks.
	Then <b>srun</b> is forked and executed again.
	The environment variable will be used by the <b>srun</b> command to restart
	execution of the tasks from the previous checkpoint.</p>

	<h3>sbatch</h3>

	<p>Several options have been added to support checkpoint restart:</p>
	<ul>
	<li><b>--checkpoint</b>: Specify the interval between periodic checkpoint
	of a batch job, in seconds</li>
	<li><b>--checkpoint-dir</b>:Specify the directory when the checkpoint image
	files of a batch job will be stored.
	The default value is the current working directory.
	Checkpoint files will be of the form <i>"<job_id>.ckpt"</i> for jobs
	and <i>"<job_id>.<step_id>.ckpt"</i> for job steps.</li>
	</li>
	</ul>

	<p>Environment variables are available for all of these options:</p>
	<ul>
	<li<b>SLURM_CHECKPOINT</b> is equivalent to <b>--checkpoint</b>:</li>
	<li><b>SLURM_CHECKPOINT_DIR</b> is equivalent to <b>--checkpoint-dir</b></li>
	</li>
	</ul>

	<h3>scontrol</h3>

	<p><b>scontrol</b> is used to initiate checkpoint/restart requests.</p>
	<ul>
	<li><b>scontrol checkpoint create <i>jobid</i> [ImageDir=<i>dir</i>]
	[MaxWait=<i>seconds</i>]</b><br>
	Requests a checkpoint on a specific job.
	For backward compatibility, if a job id is specified, all job steps of
	it are checkpointed.
	If a batch job id is specified, the entire job is checkpointed including
	the batch shell and all running tasks of all job steps.
	Upon checkpoint, the task launch command must forward the requests to
	tasks it launched.
	<ul>
	<li><b>ImageDir</b> specifies the directory in which to save the checkpoint
	image files. If specified, this takes precedence over any <b>--checkpoint-dir</b>
	option specified when the job or job step were submitted.</li>
	<li><b>MaxWait</b> specifies the maximum time permitted for a checkpoint
	request to complete. The request will be considered failed if not
	completed in this time period.</li>
	</li>
	</ul>

	<li><b>scontrol checkpoint create <i>jobid.stepid</i> [ImageDir=<i>dir</i>]
	[MaxWait=<i>seconds</i>]</b><br>
	Requests a checkpoint on a specific job step.</li>

	<li><b>scontrol checkpoint restart <i>jobid</i> [ImageDir=<i>dir</i>]
	[StickToNodes]</b><br>
	Restart a previously checkpointed batch job.
	<ul>
	<li><b>ImageDir</b> specifies the directory from which to read the checkpoint
	image files.</li>
	<li><b>StickToNodes</b> specifies that the job should be restarted on the
	same set of nodes from which it was previously checkpointed.</li>
	</ul></li>
	</ul>

	<h2>Configuration</h2>

	<p>The following SLURM configuration parameter has been added:</p>
	<ul>
	<li><b>JobCheckpointDir</b> specified the default directory for storing
	or reading job checkpoint files</li>
	</ul>

	<p class="footer"><a href="#top">top</a></p>

	<p style="text-align:center;">Last modified 11 March 2009</p>

	<!--#include virtual="footer.txt"-->