HFSS15: How to Select the Windows® HPC Job Unit Type

When the unit type is selected for a job, it also applies to all of the tasks of this job. For example, if the unit type for the job is nodes, then a user cannot schedule tasks in units of cores. It is important to select the unit type before selecting the number of units for the job or for the job tasks because the unit type affects many other choices.

If the job tasks are CPU bound, it is usually best to select a unit type of core. This allows the tasks to be distributed to as many cores as possible, making maximum of the computational resources available. This choice may also allow cores that are not currently assigned to a job to be used by other jobs.

If the job tasks are memory bound, a unit type of socket may be a better selection. Running multiple tasks on the same socket may decrease performance due to memory bus contention. Some experimentation may be needed to determine whether a socket or core unit type is more efficient.

The node unit type selection may be best if the tasks are multithreaded and effectively use the computing resources available. In these cases, there may be no benefit to scheduling multiple tasks on a node. The node unit type may also be the best selection if some system wide resource is the bottleneck. Some possible scenarios where this could occur are:

• Each task requires all of the memory available on the node, and running multiple tasks results in excessive paging or swapping

• Each task is limited by bandwidth of the disk connection

• Each task is limited by network bandwidth

References

• MSDN: ISchedulerJob::UnitType Property

• The Windows HPC Team Blog: How that Node/Socket/Core thing works