Available Partitions
Partitions are used on Swan to distinguish different
resources. You can view the partitions with the command sinfo
.
Swan:¶
Swan has a two shared public partitions available for use. The default partition batch
and the GPU enabled partition, gpu
.
When you submit a job on Swan without specifying a partition, it will automatically use the batch
Partition Name | Notes |
---|---|
batch | Default Paritition Does not have GPUs |
gpu | Shared partition with GPUs |
On Swan jobs have a maximum runtime of 7 days, can request up to 2000 cores per user, and run up to 1000 jobs.
Worker Node Configuration¶
The standard configuration of a Swan worker node is:
Configuration | Value |
---|---|
Cores | 56 |
Memory | 250 GB |
Scratch Storage | 3.5 TB |
Some Swan worker nodes are equipped with additional memory, with up to 2TB of memory available in some nodes.
GPU Enabled Worker Nodes¶
For GPU enabled worker nodes in the gpu partition, the following GPUs are available:
Description | SLURM Feature | Available Hardware |
---|---|---|
Tesla V100, with 10GbE | gpu_v100 | 1 node - 4 GPUs with 16 GB per node |
Tesla V100, with OPA | gpu_v100 | 21 nodes - 2 GPUs with 32GB per node |
Tesla V100S | gpu_v100 | 4 nodes - 2 GPUs with 32GB per node |
Tesla T4 | gpu_t4 | 12 nodes - 2 GPUs with 16GB per node |
NVIDIA A30 | gpu_a30 | 2 nodes - 4 GPUs with 24GB per node |
Additional GPUs are available in the guest_gpu
partition, but jobs running on this partition will be preemptable. Details on how the partition operates is available below in Guest Partition(s). The GPUs in this partition are listed in the partition list for Swan for the priority access partitions .
Resource requests and utilization
Please make sure your applications and software support the resources you are requesting. Many applications are only able to use a single worker node and may not scale well with large numbers of cores.
Please review our information on how many resources to request in our FAQ
For GPU monitoring and resource requests, please review our page on monitoring and optimizing GPU resources
A full list of partitions is available for Swan
SLURM Quality of Service¶
Swan has two available Quality of Service types available which help manage how the job gets scheduled. Overall limitations of maximum job wall time. CPUs, etc. are set for all jobs with the default setting (when thea "–qos=" section is omitted) and "short" jobs (described as above) on Swan. The limitations are shown in the following form.
SLURM Specification | Max Job Run Time | Max CPUs per User | Max Jobs per User | |
---|---|---|---|---|
Default | Leave blank | 7 days | 2000 | 1000 |
Short | #SBATCH --qos=short | 6 hours | 16 | 2 |
Please also note that the memory and local hard drive limits are subject to the physical limitations of the nodes, described in the resources capabilities section of the HCC Documentation and the partition sections above.
Priority for short jobs¶
To run short jobs for testing and development work, a job can specify a different quality of service (QoS). The short QoS increases a jobs priority so it will run as soon as possible.
SLURM Specification |
---|
#SBATCH --qos=short |
Limits per user for 'short' QoS
- 6 hour job run time
- 2 jobs of 16 CPUs or fewer
- No more than 256 CPUs in use for short jobs from all users
Owned/Priority Access Partitions¶
Partitions marked as owned by a group means only specific groups are allowed to submit jobs to that partition. Groups are manually added to the list allowed to submit jobs to the partition. If you are unable to submit jobs to a partition, and you feel that you should be, please contact hcc-support@unl.edu.
To submit jobs to an owned partition, use the SLURM --partition
option. Jobs
can either be submitted only to an owned partition, or to both the owned
partition and the general access queue. For example, assuming a partition
named mypartition
:
Submit only to an owned partition
#SBATCH --partition=mypartition
Submitting solely to an owned partition means jobs will start immediately until the resources on the partition are full, then queue until prior jobs finish and resources become available.
Submit to both an owned partition and general queue
#SBATCH --partition=mypartition,batch
Submitting to both an owned partition and batch
means jobs will run on both the owned
partition and the general batch queue. Jobs will start immediately until the resources
on the partition are full, then queue. Pending jobs will then start either on the owned partition
or in the general queue, wherever resources become available first
(taking into account FairShare). Unless there are specific reasons to limit jobs
to owned resources, this method is recommended to maximize job throughput.
A full list of partitions is available for Swan
Guest Partition(s)¶
The guest
partition can be used by users and groups that do not own
dedicated resources on Swan. Jobs running in the guest
partition
will run on the owned resources with Intel OPA interconnect. The jobs
are preempted when the resources are needed by the resource owners:
guest jobs will be killed and returned to the queue in a pending state
until they can be started on another node.
HCC recommends verifying job behavior will support the restart and
modifying job scripts if necessary.
To submit your job to the guest partition add the line
Submit to guest partition
#SBATCH --partition=guest
to your submit script.
Owned GPU resources may also be accessed in an opportunistic manner by
submitting to the guest_gpu
partition. Similar to guest
, jobs are
preempted when the GPU resources are needed by the owners. To submit
your job to the guest_gpu
partition, add the lines
Submit to guest_gpu partition
#SBATCH --partition=guest_gpu
#SBATCH --gres=gpu
to your SLURM script.
Preventing job restart¶
By default, jobs on the guest
partition will be restarted elsewhere when they
are preempted. To prevent preempted jobs from being restarted add the line
Prevent job restart on guest partition
#SBATCH --no-requeue
to your SLURM submit file.