Difference between revisions of "SLURM Partitions"
Line 135: | Line 135: | ||
|cms16, preempt | |cms16, preempt | ||
|CMS nodes (2 hour grace period) | |CMS nodes (2 hour grace period) | ||
|*See [https://wiki.rc.usf.edu/index.php/SLURM_Partitions# | |*See [https://wiki.rc.usf.edu/index.php/SLURM_Partitions#Preemption Preemption] description below | ||
|- | |- | ||
|mri2016 | |mri2016 | ||
Line 141: | Line 141: | ||
|mri16, preempt | |mri16, preempt | ||
|MRI nodes (2 hour grace period) | |MRI nodes (2 hour grace period) | ||
|*See [https://wiki.rc.usf.edu/index.php/SLURM_Partitions# | |*See [https://wiki.rc.usf.edu/index.php/SLURM_Partitions#Preemption Preemption] description below | ||
|- | |- | ||
|himem | |himem |
Revision as of 20:21, 9 January 2017
SLURM Partitions
Dispatching
It is really no longer necessary to discuss queues in the traditional sense. In the past, we would create queues based on pools of hardware resources. If a user wanted to utilize a particular hardware resource, he or she would request the appropriate queue. Most times, however, what the user wants and what is best for the user or what is best for all users are not necessarily the same. Allowing individuals to dictate where their jobs will run will inevitably lead to throughput problems since it would be unreasonable to expect the users to understand the complete state and behavior of the scheduler.
Below is a general description of how jobs make their way through the queue. Please see Scheduling and Dispatch Policy for more information.
When a user submits a job to a specific partiton, the scheduler determines if the requested hardware/time requirements of the job (see Using Features) match up with the resources the partition provides. If it does, the job is executed if there are available resources. If there are no available resources, the job will be held until the next scheduler iteration, to see if resources have become available.
“Available resources” include processors and memory. Processors generally match up to the number of slots in a given queue while memory is defined as a complex value which may not be so obvious to query. If your job is waiting in the qw
state, it is likely that either the slots requested or the memory requested are beyond what the system can provide at that particular point in time.
Partition Node Sets
The following node sets are available:
Memory | CPU | Cores | Interconnect | Nodes | Slots | GPUs | Complex Flags | Location |
24GB | Opteron 2384 | 12 | 4x DDR IB | 32 | 384 | n/a | ib_ddr, ib_psm, tpa, sse4, sse4a, cpu_amd, opteron_2384 | Tampa |
24GB | Xeon E5649 | 12 | 4x QDR IB | 107 | 1284 | 8 | ib_qdr, ib_psm, sse4, sse41, sse42, cpu_xeon, xeon_E5649 | Tampa |
24GB | Xeon E5-2630 | 12 | 4x QDR IB | 67 | 804 | n/a | ib_qdr, ib_psm, sse41, sse42, avx, cpu_xeon, xeon_E52630 | Tampa |
24GB | Xeon E5649 | 12 | 4x QDR IB | 14 | 168 | n/a | ib_qdr, ib_ofa, sse41, sse42, avx, cpu_xeon, xeon_E5649 | Tampa |
32GB | Xeon E5-2670 | 16 | 4x QDR IB | 129 | 2064 | 40 | ib_qdr, ib_psm, sse4, sse41, sse42, avx, cpu_xeon, xeon_E52670, gpu_K20 | Tampa |
512GB | Xeon E5-2650 | 20 | 4x QDR IB | 3 | 60 | n/a | ib_qdr, ib_psm, tpa, sse4_1, sse4_2, avx, avx2, gpfs, cpu_xeon, xeon_E52650, mem_512G | Tampa |
Total | 4764 | 48 |
Partition Layout
The node sets are associated with the following queues:
Queue Name | Max Runtime | QOS' Required | Description/Preempt Grace Period | Notes |
circe | infinite | none | default general-purpose queue | |
rc2016 | infinite | none | general-purpose queue with latest hardware | |
gpfsgpu | infinite | none | CUDA GPU nodes | |
cuda | infinite | none | CUDA GPU nodes | |
cms2016 | infinite | cms16, preempt | CMS nodes (2 hour grace period) | *See Preemption description below |
mri2016 | infinite | mri16, preempt | MRI nodes (2 hour grace period) | *See Preemption description below |
himem | 1 week | memaccess | large memory job queue (>= 24 GB) | To request access, email rc-help@usf.edu |
Preemption
Some hardware on CIRCE is provided by research contributors. This hardware is available for use by all CIRCE users by specifying the partition and the "preempt" QOS (example: sbatch --partition=mri2016 --qos=preempt ./submit-script.sh). The caveat however is that as this is contributor hardware, non-contributor jobs running on this partition are subject to preemption.
There is a partition-specified grace period (listed above, typically 2 hours) before a contributor’s job(s) will cancel the non-contributors job(s). This means that any user taking advantage of the hardware should have some kind of check-pointing enabled, so that interrupted jobs can be re-submitted without needing to start over.