HFSS15: Specifying the Number of Compute Resource Units for HPC Jobs

The number of resource units (irrespective of the type of the unit) must always be greater than or equal to the Number of engines. Apart from this requirement, you need to set correct values for Compute resource granularity, Number of compute resource units, Node exclusivity and Job description file that ensure that the job has sufficient resources. Following are some examples:

For a pure DSO job that does not use much memory:

• Set Compute resource granularity to core or socket (see Selecting Computation Resource Units (Job Unit Type)).

• Set Number of compute resource units to the 'Number of engines'.

For a job that is memory intensive. Run a fixed number of engines on each node:

• Set Compute resource granularity to core or socket (see How to Select the HPC Job Unit Type).

• Set Node exclusivity to true.

• Set Number of compute resource units to the number of units that are for exclusive use by this job.

• Specify an appropriate node list or the node groups via the Job description file.

Example: Suppose the job is run on nodes with 32GB of memory and 8 cores per node. Suppose each engine requires 8GB memory and that 4 engines can run on each machine. Suppose the Number of engines specified for this job is 10. Since four engines are run per node, we need three nodes to run 10 engines. So, 'Number of compute resource units' should be set to 24 cores (3 nodes times 8 cores). The Number of cores per engine should be set to 2.The Number of cores should be set to 8 (or a smaller number, if using 8 cores slows down the analysis)

For a job that is memory intensive. Run one engine per node:

• Set Compute resource granularity to Node

• Set Number of compute resource units to Number of engines.

• Specify appropriate node list or node groups via the Job description file.

Example: Suppose the job is run on nodes with 16GB of memory and 4 cores per node. Suppose each engine requires 16GB memory, that is, 1 engine can run on each machine. Suppose the Number of engines specified for this job is 10. So, the Number of compute resource units should be set to 10 nodes. The Number of cores per engine should be set to 4 (or a smaller number, if using 4 cores slows down the analysis). The Number of cores should be set to 4 (or a smaller number, if using 4 cores slows down the analysis)