Skip to content

Partitions

Midway compute nodes are organized into "partitions", collections of compute nodes with similar characteristics, such as hardware configuration or ownership. When a user submits a job to a partition (via Slurm flag --partition=<partition>), the job is allocated by the Slurm scheduler to any idle compute node within that partition. If a user wants to submit a job to a particular compute node, this can be requested by adding --nodelist=<compute_node_ID>. Most users will submit jobs to a 'shared' partition: a partition accessible to all Midway users. See the Slurm Partitions page for additional information about the Midway partitions.

How Do I Get a Full List of Partitions?

To get a full list of partitions available On Midway2 or Midway3, log in to the system and type in the terminal:

sinfo -o "%20P %5D %14F %4c %8G %8z %26f %N"
The typical output will include:

Column Description
AVAIL_FEATURES Available features such as CPUs, GPUs, and internode interfaces
NODELIST IDs of the compute nodes within the given partition
NODES(A/I/O/T) Number of nodes by state in the format "allocated/idle/other/total"
S:C:T Number of sockets, cores, and threads

If a user wants to submit their job to the particular compute node, this can be requested by adding the Slurm flag --nodelist=<compute_node_ID>. Compute nodes that differ in available features can be allocated by setting an additional constraint --constraint=<compute_node_feature>; for example, --constraint=v100 will allocate the job to the compute node with NVIDIA V100 GPUs.

You can also check the state of nodes in a given partition, for example, caslake, by running the following command:

nodestatus caslake

What Partitions I Can Use?

To submit jobs to a partition, ensure that your account is accepted by the partition. Begin by determining the accounts you belong to. Then set the variable ACCOUNT to the account you would like to check (usually in the format pi-), and then retrieve the list of partitions using:

groups

export ACCOUNT=<account_name>
sacctmgr list assoc account=$ACCOUNT -n format=Partition | sort -u

export ACCOUNT=<account_name>
scontrol show partition | grep -B 1 -P "(?=.*AllowAccounts)(?=.*$ACCOUNT) | (?=.*AllowAccounts=ALL)" | grep Partition | sed 's/.*=//'

Configurations

All Midway users can submit jobs to any of the shared partitions. Beagle3 users, in addition to shared partitions, have access to Beagle3 partitions too.

Parameters are shown per node.

Partition
# of Nodes
# of Cores/node
CPU Type # of GPUs/node GPU Type Memory/node
Nodelist
broadwl 323 28 e5-2680v4 None None 64 GB vary
broadwl-lc 14 28 e5-2680v4 None None 64 GB midway2-[0203-0216]
bigmem2 5 28 e5-2680v4 None None 512 GB midway2-bigmem[01-04,06]
gpu2 6 28 e5-2680v4 4 k80 128 GB midway2-gpu[01-06]
Partition
# of Nodes
# of Cores/node
CPU Type # of GPUs/node GPU Type Memory/node
Nodelist
dali 29 40 gold-6148 None None 96 GB vary
Partition # of Nodes # of Cores/node CPU Type # of GPUs/node GPU Type Memory/node
Nodelist
caslake 184 48 gold-6248r None None 192 GB vary
bigmem 1 48 gold-6248r None None 768 GB midway3-0317
bigmem 1 48 gold-6248r None None 1536 GB midway3-0318
amd-hm 1 128 epyc-7702 None None 2048 GB midway3-0541
amd 40 128 epyc-7702 None None 256 GB midway3-0[501-539,549]
gpu 1 48 gold-6248r 4 a100 384 GB midway3-0294
gpu 5 48 gold-6248r 4 v100 192 GB midway3-[0277-0281]
gpu 5 48 gold-6248r 4 rtx6000 192 GB midway3-[0282-0286]
Partition # of Nodes # of Cores/node CPU Type # of GPUs/node GPU Type Memory/node Nodelist
beagle3 22 32 gold-6346 4 a40 256 GB beagle3-[0001-0022]
beagle3 22 32 gold-6346 4 a100 256 GB beagle3-[0023-0044]
beagle3 4 32 gold-6346 None None 512 GB beagle3-bigmem[1-4]
Partition # of Nodes # of Cores/node CPU Type # of GPUs/node GPU Type Memory/node
ssd 18 48 gold-6248r None None 192 GB
ssd-gpu 1 32 gold-6346 4 a100 256 GB
Partition # of Nodes # of Cores/node CPU Type # of GPUs/node GPU Type Memory/node
kicp 6 48 gold-6248r None None 192 GB
kicp-gpu 1 32 gold-5218 4 v100 192 GB

Partition QoS

This table details the job limits of each partition, set via a Quality of Service (QoS) accessible via

rcchelp qos

Partition Max Nodes/User Max CPUs/User Max Jobs/User Max Wall Time
broadwl 100 2800 100 36 H
gpu2 None None 10 36 H
bigmem2 None 112 5 36 H
Partition Max Nodes/User Max CPUs/User Max Jobs/User Max Wall Time
caslake 100 4800 1000 36 H
gpu None None 12 36 H
bigmem 96 192 10 36 H
amd 64 128 200 36 H
Partition Max Nodes/User Max CPUs/User Max Jobs/User Max Wall Time AllowAccount QoS
ssd 5 240 N/A 36 H ssd ssd
ssd 5 240 N/A 36 H ssd-stu ssd-stu
ssd-gpu N/A N/A N/A 36 H ssd ssd
ssd-gpu N/A N/A N/A 36 H ssd-stu ssd-stu
Partition AllowAccount QoS Max Wall Time
kicp kicp kicp 48 H
kicp-gpu kicp kicp 48 H

You may apply for a special allocation if your research requires a temporary exception to a particular limit. Special allocations are evaluated on an individual basis and may or may not be granted.

Private Partitions (CPP)

These partitions are typically associated with a PI or group of PIs. They can be shared with other PIs or researchers across the University of Chicago and with external collaborators with a Midway account.

  • General Users: Check with your PI for more information.
  • PIs: Private partitions can be purchased via Cluster Partnership Program to accommodate the needs of your research group. QOS can be tuned at any time by submitting a request.

Note

QoS for private and institutional partitions can be changed upon the owner's request.