Gang Scheduling

Slurm supports timesliced gang scheduling in which two or more jobs are allocated to the same resources and these jobs are alternately suspended to let one job at a time have dedicated access to the resources for a configured period of time. Slurm also supports preemptive priority job scheduling in which a a higher priority job can preempt a lower priority one until the higher priority job completes. See the Preemption document for more information.

A workload manager that supports timeslicing can improve responsiveness and utilization by allowing more jobs to begin running sooner. Shorter-running jobs no longer have to wait in a queue behind longer-running jobs. Instead they can be run "in parallel" with the longer-running jobs, which will allow them to start and finish quicker. Throughput is also improved because overcommitting the resources provides opportunities for "local backfilling" to occur (see example below).

The gang scheduling logic monitors each of the partitions in Slurm. If a new job has been allocated to resources in a partition that have already been allocated to an existing job, then the plugin will suspend the new job until the configured SchedulerTimeslice interval has elapsed. Then it will suspend the running job and let the new job make use of the resources for a SchedulerTimeslice interval. This will continue until one of the jobs terminates.

Configuration

There are several important configuration parameters relating to gang scheduling:

  • SelectType: The Slurm gang scheduler supports nodes allocated by the select/linear plugin, socket/core/CPU resources allocated by the select/cons_res plugin or select/cray for Cray systems without ALPS.
  • SelectTypeParameter: Since resources will be getting overallocated with jobs, the resource selection plugin should be configured to track the amount of memory used by each job to ensure that memory page swapping does not occur. When select/linear is chosen, we recommend setting SelectTypeParameter=CR_Memory. When select/cons_res is chosen, we recommend including Memory as a resource (ex. SelectTypeParameter=CR_Core_Memory).
  • DefMemPerCPU: Since job requests may not explicitly specify a memory requirement, we also recommend configuring DefMemPerCPU (default memory per allocated CPU) or DefMemPerNode (default memory per allocated node). It may also be desirable to configure MaxMemPerCPU (maximum memory per allocated CPU) or MaxMemPerNode (maximum memory per allocated node) in slurm.conf. Users can use the --mem or --mem-per-cpu option at job submission time to specify their memory requirements. Note that in order to gang schedule jobs, all jobs must be able to fit into memory at the same time.
  • JobAcctGatherType and JobAcctGatherFrequency: If you wish to enforce memory limits, either that task/cgroup must be configured to limit each job's memory use or accounting must be enabled using the JobAcctGatherType and JobAcctGatherFrequency parameters. If accounting is enabled and a job exceeds its configured memory limits, it will be canceled in order to prevent it from adversely effecting other jobs sharing the same resources.
  • PreemptMode: set the GANG option. Additional options may be specified to enable job preemption in addition to gang scheduling.
  • SchedulerTimeSlice: The default timeslice interval is 30 seconds. To change this duration, set SchedulerTimeSlice to the desired interval (in seconds) in slurm.conf. For example, to set the timeslice interval to one minute, set SchedulerTimeSlice=60. Short values can increase the overhead of gang scheduling.
  • OverSubscribe: Configure the partition's OverSubscribe setting to FORCE for all partitions in which timeslicing is to take place. The FORCE option supports an additional parameter that controls how many jobs can share a compute resource (FORCE[:max_share]). By default the max_share value is 4. To allow up to 6 jobs from this partition to be allocated to a common resource, set OverSubscribe=FORCE:6. To only let 2 jobs timeslice on the same resources, set OverSubscribe=FORCE:2. NOTE: Gang scheduling is performed independently for each partition, so configuring partitions with overlapping nodes and gang scheduling is generally not recommended.

In order to enable gang scheduling after making the configuration changes described above, restart Slurm if it is already running. Any change to the plugin settings in Slurm requires a full restart of the daemons. If you just change the partition OverSubscribe setting, this can be updated with scontrol reconfig.

Timeslicer Design and Operation

When enabled, the gang scheduler keeps track of the resources allocated to all jobs. For each partition an "active bitmap" is maintained that tracks all concurrently running jobs in the Slurm cluster. Each time a new job is allocated to resources in a partition, the gang scheduler compares these newly allocated resources with the resources already maintained in the "active bitmap". If these two sets of resources are disjoint then the new job is added to the "active bitmap". If these two sets of resources overlap then the new job is suspended. All jobs are tracked in a per-partition job queue within the gang scheduler logic.

A separate timeslicer thread is spawned by the gang scheduler on startup. This thread sleeps for the configured SchedulerTimeSlice interval. When it wakes up, it checks each partition for suspended jobs. If suspended jobs are found then the timeslicer thread moves all running jobs to the end of the job queue. It then reconstructs the "active bitmap" for this partition beginning with the suspended job that has waited the longest to run (this will be the first suspended job in the run queue). Each following job is then compared with the new "active bitmap", and if the job can be run concurrently with the other "active" jobs then the job is added. Once this is complete then the timeslicer thread suspends any currently running jobs that are no longer part of the "active bitmap", and resumes jobs that are new to the "active bitmap".

This timeslicer thread algorithm for rotating jobs is designed to prevent jobs from starving (remaining in the suspended state indefinitely) and to be as fair as possible in the distribution of runtime while still keeping all of the resources as busy as possible.

The gang scheduler suspends jobs via the same internal functions that support scontrol suspend and scontrol resume. A good way to observe the operation of the timeslicer is by running squeue -i<time> in a terminal window where time is set equal to SchedulerTimeSlice.

A Simple Example

The following example is configured with select/linear and OverSubscribe=FORCE. This example takes place on a small cluster of 5 nodes:

[user@n16 load]$ sinfo
PARTITION AVAIL  TIMELIMIT NODES  STATE NODELIST
active*      up   infinite     5   idle n[12-16]

Here are the Scheduler settings (excerpt of output):

[user@n16 load]$ scontrol show config
...
PreemptMode             = GANG
...
SchedulerTimeSlice      = 30
SchedulerType           = sched/builtin
...

The myload script launches a simple load-generating app that runs for the given number of seconds. Submit myload to run on all nodes:

[user@n16 load]$ sbatch -N5 ./myload 300
sbatch: Submitted batch job 3

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    3    active  myload  user     0:05     5 n[12-16]

Submit it again and watch the gang scheduler suspend it:

[user@n16 load]$ sbatch -N5 ./myload 300
sbatch: Submitted batch job 4

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    3    active  myload  user  R  0:13     5 n[12-16]
    4    active  myload  user  S  0:00     5 n[12-16]

After 30 seconds the gang scheduler swaps jobs, and now job 4 is the active one:

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    4    active  myload  user  R  0:08     5 n[12-16]
    3    active  myload  user  S  0:41     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    4    active  myload  user  R  0:21     5 n[12-16]
    3    active  myload  user  S  0:41     5 n[12-16]

After another 30 seconds the gang scheduler sets job 3 running again:

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    3    active  myload  user  R  0:50     5 n[12-16]
    4    active  myload  user  S  0:30     5 n[12-16]

A possible side effect of timeslicing: Note that jobs that are immediately suspended may cause their srun commands to produce the following output:

[user@n16 load]$ cat slurm-4.out
srun: Job step creation temporarily disabled, retrying
srun: Job step creation still disabled, retrying
srun: Job step creation still disabled, retrying
srun: Job step creation still disabled, retrying
srun: Job step created

This occurs because srun is attempting to launch a jobstep in an allocation that has been suspended. The srun process will continue in a retry loop to launch the jobstep until the allocation has been resumed and the jobstep can be launched.

When the gang scheduler is enabled, this type of output in the user jobs should be considered benign.

More examples

The following example shows how the timeslicer algorithm keeps the resources busy. Job 10 runs continually, while jobs 9 and 11 are timesliced:

[user@n16 load]$ sbatch -N3 ./myload 300
sbatch: Submitted batch job 9

[user@n16 load]$ sbatch -N2 ./myload 300
sbatch: Submitted batch job 10

[user@n16 load]$ sbatch -N3 ./myload 300
sbatch: Submitted batch job 11

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    9    active  myload  user  R  0:11     3 n[12-14]
   10    active  myload  user  R  0:08     2 n[15-16]
   11    active  myload  user  S  0:00     3 n[12-14]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   10    active  myload  user  R  0:50     2 n[15-16]
   11    active  myload  user  R  0:12     3 n[12-14]
    9    active  myload  user  S  0:41     3 n[12-14]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   10    active  myload  user  R  1:04     2 n[15-16]
   11    active  myload  user  R  0:26     3 n[12-14]
    9    active  myload  user  S  0:41     3 n[12-14]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
    9    active  myload  user  R  0:46     3 n[12-14]
   10    active  myload  user  R  1:13     2 n[15-16]
   11    active  myload  user  S  0:30     3 n[12-14]

The next example displays "local backfilling":

[user@n16 load]$ sbatch -N3 ./myload 300
sbatch: Submitted batch job 12

[user@n16 load]$ sbatch -N5 ./myload 300
sbatch: Submitted batch job 13

[user@n16 load]$ sbatch -N2 ./myload 300
sbatch: Submitted batch job 14

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   12    active  myload  user  R  0:14     3 n[12-14]
   14    active  myload  user  R  0:06     2 n[15-16]
   13    active  myload  user  S  0:00     5 n[12-16]

Without timeslicing and without the backfill scheduler enabled, job 14 has to wait for job 13 to finish.

This is called "local" backfilling because the backfilling only occurs with jobs close enough in the queue to get allocated by the scheduler as part of oversubscribing the resources. Recall that the number of jobs that can overcommit a resource is controlled by the OverSubscribe=FORCE:max_share value, so this value effectively controls the scope of "local backfilling".

Normal backfill algorithms check all jobs in the wait queue.

Consumable Resource Examples

The following two examples illustrate the primary difference between CR_CPU and CR_Core when consumable resource selection is enabled (select/cons_res).

When CR_CPU (or CR_CPU_Memory) is configured then the selector treats the CPUs as simple, interchangeable computing resources unless task affinity is enabled. However when task affinity is enabled with CR_CPU or CR_Core (or CR_Core_Memory) is enabled, the selector treats the CPUs as individual resources that are specifically allocated to jobs. This subtle difference is highlighted when timeslicing is enabled.

In both examples 6 jobs are submitted. Each job requests 2 CPUs per node, and all of the nodes contain two quad-core processors. The timeslicer will initially let the first 4 jobs run and suspend the last 2 jobs. The manner in which these jobs are timesliced depends upon the configured SelectTypeParameter.

In the first example CR_Core_Memory is configured. Note that jobs 46 and 47 don't ever get suspended. This is because they are not sharing their cores with any other job. Jobs 48 and 49 were allocated to the same cores as jobs 44 and 45. The timeslicer recognizes this and timeslices only those jobs:

[user@n16 load]$ sinfo
PARTITION AVAIL  TIMELIMIT NODES  STATE NODELIST
active*      up   infinite     5   idle n[12-16]

[user@n16 load]$ scontrol show config | grep Select
SelectType              = select/cons_res
SelectTypeParameters    = CR_CORE_MEMORY

[user@n16 load]$ sinfo -o "%20N %5D %5c %5z"
NODELIST             NODES CPUS  S:C:T
n[12-16]             5     8     2:4:1

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 44

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 45

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 46

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 47

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 48

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 49

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   44    active  myload  user  R  0:09     5 n[12-16]
   45    active  myload  user  R  0:08     5 n[12-16]
   46    active  myload  user  R  0:08     5 n[12-16]
   47    active  myload  user  R  0:07     5 n[12-16]
   48    active  myload  user  S  0:00     5 n[12-16]
   49    active  myload  user  S  0:00     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   46    active  myload  user  R  0:49     5 n[12-16]
   47    active  myload  user  R  0:48     5 n[12-16]
   48    active  myload  user  R  0:06     5 n[12-16]
   49    active  myload  user  R  0:06     5 n[12-16]
   44    active  myload  user  S  0:44     5 n[12-16]
   45    active  myload  user  S  0:43     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   44    active  myload  user  R  1:23     5 n[12-16]
   45    active  myload  user  R  1:22     5 n[12-16]
   46    active  myload  user  R  2:22     5 n[12-16]
   47    active  myload  user  R  2:21     5 n[12-16]
   48    active  myload  user  S  1:00     5 n[12-16]
   49    active  myload  user  S  1:00     5 n[12-16]

Note the runtime of all 6 jobs in the output of the last squeue command. Jobs 46 and 47 have been running continuously, while jobs 44 and 45 are splitting their runtime with jobs 48 and 49.

The next example has CR_CPU_Memory configured and the same 6 jobs are submitted. Here the selector and the timeslicer treat the CPUs as countable resources which results in all 6 jobs sharing time on the CPUs:

[user@n16 load]$ sinfo
PARTITION AVAIL  TIMELIMIT NODES  STATE NODELIST
active*      up   infinite     5   idle n[12-16]

[user@n16 load]$ scontrol show config | grep Select
SelectType              = select/cons_res
SelectTypeParameters    = CR_CPU_MEMORY

[user@n16 load]$ sinfo -o "%20N %5D %5c %5z"
NODELIST             NODES CPUS  S:C:T
n[12-16]             5     8     2:4:1

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 51

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 52

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 53

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 54

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 55

[user@n16 load]$ sbatch -n10 -N5 ./myload 300
sbatch: Submitted batch job 56

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   51    active  myload  user  R  0:11     5 n[12-16]
   52    active  myload  user  R  0:11     5 n[12-16]
   53    active  myload  user  R  0:10     5 n[12-16]
   54    active  myload  user  R  0:09     5 n[12-16]
   55    active  myload  user  S  0:00     5 n[12-16]
   56    active  myload  user  S  0:00     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   51    active  myload  user  R  1:09     5 n[12-16]
   52    active  myload  user  R  1:09     5 n[12-16]
   55    active  myload  user  R  0:23     5 n[12-16]
   56    active  myload  user  R  0:23     5 n[12-16]
   53    active  myload  user  S  0:45     5 n[12-16]
   54    active  myload  user  S  0:44     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   53    active  myload  user  R  0:55     5 n[12-16]
   54    active  myload  user  R  0:54     5 n[12-16]
   55    active  myload  user  R  0:40     5 n[12-16]
   56    active  myload  user  R  0:40     5 n[12-16]
   51    active  myload  user  S  1:16     5 n[12-16]
   52    active  myload  user  S  1:16     5 n[12-16]

[user@n16 load]$ squeue
JOBID PARTITION    NAME  USER ST  TIME NODES NODELIST
   51    active  myload  user  R  3:18     5 n[12-16]
   52    active  myload  user  R  3:18     5 n[12-16]
   53    active  myload  user  R  3:17     5 n[12-16]
   54    active  myload  user  R  3:16     5 n[12-16]
   55    active  myload  user  S  3:00     5 n[12-16]
   56    active  myload  user  S  3:00     5 n[12-16]

Note that the runtime of all 6 jobs is roughly equal. Jobs 51-54 ran first so they're slightly ahead, but so far all jobs have run for at least 3 minutes.

At the core level this means that Slurm relies on the Linux kernel to move jobs around on the cores to maximize performance. This is different than when CR_Core_Memory was configured and the jobs would effectively remain "pinned" to their specific cores for the duration of the job. Note that CR_Core_Memory supports CPU binding, while CR_CPU_Memory does not.

Note that manually suspending a job (i.e. "scontrol suspend ...") releases its CPUs for allocation to other jobs. Resuming a previously suspended job may result in multiple jobs being allocated the same CPUs, which could trigger gang scheduling of jobs. Use of the scancel command to send SIGSTOP and SIGCONT signals would stop a job without releasing its CPUs for allocaiton to other jobs and would be a preferable mechanism in many cases.

Last modified 31 March 2016