Sharing Consumable Resources

CPU Management

(Disclaimer: In this "CPU Management" section, the term "consumable resource" does not include memory. The management of memory as a consumable resource is discussed in it's own section below.)

The per-partition OverSubscribe setting applies to the entity being selected for scheduling:

  • When the default select/linear plugin is enabled, the per-partition OverSubscribe setting controls whether or not the nodes are shared among jobs.

  • When the select/cons_res plugin is enabled, the per-partition OverSubscribe setting controls whether or not the configured consumable resources are shared among jobs. When a consumable resource such as a core, socket, or CPU is shared, it means that more than one job can be assigned to it.

The following table describes this new functionality in more detail:

Selection Setting Per-partition OverSubscribe Setting Resulting Behavior
SelectType=select/linear OverSubscribe=NO Whole nodes are allocated to jobs. No node will run more than one job per partition/queue.
OverSubscribe=YES By default same as OverSubscribe=NO. Nodes allocated to a job may be shared with other jobs if each job allows sharing via the srun --oversubscribe option.
OverSubscribe=FORCE Each whole node can be allocated to multiple jobs up to the count specified per partition/queue (default 4 jobs per node)
SelectType=select/cons_res
Plus one of the following:
SelectTypeParameters=CR_Core
SelectTypeParameters=CR_Core_Memory
OverSubscribe=NO Cores are allocated to jobs. No core will run more than one job per partition/queue.
OverSubscribe=YES By default same as OverSubscribe=NO. Cores allocated to a job may be shared with other jobs if each job allows sharing via the srun --oversubscribe option.
OverSubscribe=FORCE Each core can be allocated to multple jobs up to the count specified per partition/queue (default 4 jobs per core).
SelectType=select/cons_res
Plus one of the following:
SelectTypeParameters=CR_CPU
SelectTypeParameters=CR_CPU_Memory
OverSubscribe=NO CPUs are allocated to jobs. No CPU will run more than one job per partition/queue.
OverSubscribe=YES By default same as OverSubscribe=NO. CPUs allocated to a job may be shared with other jobs if each job allows sharing via the srun --oversubscribe option.
OverSubscribe=FORCE Each CPU can be allocated to multple jobs up to the count specified per partition/queue (default 4 jobs per CPU).
SelectType=select/cons_res
Plus one of the following:
SelectTypeParameters=CR_Socket
SelectTypeParameters=CR_Socket_Memory
OverSubscribe=NO Sockets are allocated to jobs. No Socket will run more than one job per partition/queue.
OverSubscribe=YES By default same as OverSubscribe=NO. Sockets allocated to a job may be shared with other jobs if each job allows sharing via the srun --oversubscribe option.
OverSubscribe=FORCE Each socket can be allocated to multple jobs up to the count specified per partition/queue (default 4 jobs per socket).

When OverSubscribe=FORCE is configured, the consumable resources are scheduled for jobs using a least-loaded algorithm. Thus, idle CPUs|cores|sockets will be allocated to a job before busy ones, and CPUs|cores|sockets running one job will be allocated to a job before ones running two or more jobs. This is the same approach that the select/linear plugin uses when allocating "shared" nodes.

Note that the granularity of the "least-loaded" algorithm is what distinguishes the two selection plugins (cons_res and linear) when OverSubscribe=FORCE is configured. With the select/cons_res plugin enabled, the CPUs of a node are not overcommitted until all of the rest of the CPUs are overcommitted on the other nodes. Thus if one job allocates half of the CPUs on a node and then a second job is submitted that requires more than half of the CPUs, the select/cons_res plugin will attempt to place this new job on other busy nodes that have more than half of the CPUs available for use. The select/linear plugin simply counts jobs on nodes, and does not track the CPU usage on each node.

This new sharing functionality in the select/cons_res plugin also supports the new OverSubscribe=FORCE:<num> syntax. If OverSubscribe=FORCE:3 is configured with select/cons_res and CR_Core or CR_Core_Memory, then the select/cons_res plugin will run up to 3 jobs on each core of each node in the partition. If CR_Socket or CR_Socket_Memory is configured, then the select/cons_res plugin will run up to 3 jobs on each socket of each node in the partition.

Nodes in Multiple Partitions

Slurm has supported configuring nodes in more than one partition since version 0.7.0. The following table describes how nodes configured in two partitions with different OverSubscribe settings will be allocated to jobs. Note that "shared" jobs are jobs that are submitted to partitions configured with OverSubscribe=FORCE or with OverSubscribe=YES and the job requested sharing with the srun --oversubscribe option. Conversely, "non-shared" jobs are jobs that are submitted to partitions configured with OverSubscribe=NO or OverSubscribe=YES and the job did not request sharable resources.

 First job "sharable"First job not "sharable"
Second job "sharable"Both jobs can run on the same nodes and may share resourcesJobs do not run on the same nodes
Second job not "sharable"Jobs do not run on the same nodes Jobs can run on the same nodes but will not share resources

The next table contains several scenarios with the select/cons_res plugin enabled to further clarify how a node is used when it is configured in more than one partition and the partitions have different "OverSubscribe" policies:

Slurm configuration Resulting Behavior
Two OverSubscribe=NO partitions assigned the same set of nodes Jobs from either partition will be assigned to all available consumable resources. No consumable resource will be shared. One node could have 2 jobs running on it, and each job could be from a different partition.
Two partitions assigned the same set of nodes: one partition is OverSubscribe=FORCE, and the other is OverSubscribe=NO A node will only run jobs from one partition at a time. If a node is running jobs from the OverSubscribe=NO partition, then none of it's consumable resources will be shared. If a node is running jobs from the OverSubscribe=FORCE partition, then it's consumable resources can be shared.
Two OverSubscribe=FORCE partitions assigned the same set of nodes Jobs from either partition will be assigned consumable resources. All consumable resources can be shared. One node could have 2 jobs running on it, and each job could be from a different partition.
Two partitions assigned the same set of nodes: one partition is OverSubscribe=FORCE:3, and the other is OverSubscribe=FORCE:5 Generally the same behavior as above. However no consumable resource will ever run more than 3 jobs from the first partition, and no consumable resource will ever run more than 5 jobs from the second partition. A consumable resource could have up to 8 jobs running on it at one time.

Note that the "mixed shared setting" configuration (row #2 above) introduces the possibility of starvation between jobs in each partition. If a set of nodes are running jobs from the OverSubscribe=NO partition, then these nodes will continue to only be available to jobs from that partition, even if jobs submitted to the OverSubscribe=FORCE partition have a higher priority. This works in reverse also, and in fact it's easier for jobs from the OverSubscribe=FORCE partition to hold onto the nodes longer because the consumable resource "sharing" provides more resource availability for new jobs to begin running "on top of" the existing jobs. This happens with the select/linear plugin also, so it's not specific to the select/cons_res plugin.

Memory Management

The management of memory as a consumable resource remains unchanged and can be used to prevent oversubscription of memory, which would result in having memory pages swapped out and severely degraded performance.

Selection Setting Resulting Behavior
SelectType=select/linear Memory allocation is not tracked. Jobs are allocated to nodes without considering if there is enough free memory. Swapping could occur!
SelectType=select/linear plus
SelectTypeParameters=CR_Memory
Memory allocation is tracked. Nodes that do not have enough available memory to meet the jobs memory requirement will not be allocated to the job.
SelectType=select/cons_res
Plus one of the following:
SelectTypeParameters=CR_Core
SelectTypeParameters=CR_CPU
SelectTypeParameters=CR_Socket
Memory allocation is not tracked. Jobs are allocated to consumable resources without considering if there is enough free memory. Swapping could occur!
SelectType=select/cons_res
Plus one of the following:
SelectTypeParameters=CR_Memory
SelectTypeParameters=CR_Core_Memory
SelectTypeParameters=CR_CPU_Memory
SelectTypeParameters=CR_Socket_Memory
Memory allocation for all jobs are tracked. Nodes that do not have enough available memory to meet the jobs memory requirement will not be allocated to the job.

Users can specify their job's memory requirements one of two ways. The srun --mem=<num> option can be used to specify the jobs memory requirement on a per allocated node basis. This option is recommended for use with the select/linear plugin, which allocates whole nodes to jobs. The srun --mem-per-cpu=<num> option can be used to specify the jobs memory requirement on a per allocated CPU basis. This is recommended for use with the select/cons_res plugin which can allocate individual CPUs to jobs.

Default and maximum values for memory on a per node or per CPU basis can be configured by the system administrator using the following slurm.conf options: DefMemPerCPU, DefMemPerNode, MaxMemPerCPU and MaxMemPerNode. Users can use the --mem or --mem-per-cpu option at job submission time to override the default value, but they cannot exceed the maximum value.

Enforcement of a jobs memory allocation is performed by setting the "maximum data segment size" and the "maximum virtual memory size" system limits to the appropriate values before launching the tasks. Enforcement is also managed by the accounting plugin, which periodically gathers data about running jobs. Set JobAcctGather and JobAcctFrequency to values suitable for your system.

Last modified 31 March 2016