Requesting specific node types

While the HEC’s compute nodes are all similar in their underlying architecture, different eras of compute nodes are of very different sizes. For the purposes of parallel jobs which request multiple nodes, it’s important to ensure that each node used is of broadly the same type. There may also be some circumstances where a particular application may perform better on one type of node than another. Jobs can request that they be restricted to a specific type of node using the job resource request node_type by using the following syntax in a job script:

#SBATCH -C node_type=type

Where type is one of the following:

Type

# cores

Processor

Memory

Interconnect

10Geth64G

16

Intel Ivy Bridge, Haswell or Broadwell

64G

10Gb low latency

10GethAny

16

Intel Ivy Bridge, Haswell or Broadwell

64G

10Gb low latency

40core192G

40

Intel Skylake or Cascade Lake

192G

25Gb low latency

64core256G

64

Intel Ice Lake

256G

25Gb low latency

Note

The 10GethAny node type exists only to provide backwards compatibilty - it used to refer to both the existing 10Geth64G and the now retired 10Geth128G nodes.

The scheduler will handle node_type requests in an intelligent way, depending on job type:

For serial jobs:

  • If no node_type is specified, the job is free to launch on any type of compute node. The node_type should typically not be used for serial jobs.

For parallel jobs making an exclusive resource request (ie, requesting one or more whole compute nodes):

  • If no node_type is specified, the job will be run as though node_type=10Geth64G had been specified.

  • If node_type=10Geth64G or node_type=10GethAny is specified, and the memory resource request will be set to 3.95G per core (for a total of 63.2G per node).

  • If node_type=40core192G is specified, the job will be eligible to run on compute nodes of type 40core192G with a memory resource request of 4.75G per core (for a total of 190G per node).

  • If node_type=64core256G is specified, the job will be eligible to run on compute nodes of type 64core256G with a memory resource request of 3.95G per core (for a total of 252.8G per node).

For parallel jobs that do not make an exclusive request (ie, OpenMP jobs requesting a number of cores all on the same compute node):

  • If no node_type is specified the job will be eligible to run on any node that can suport the number of cores and amount of memory requested