Checking compute node performance

Before calculating coefficient data for each compute node it is necessary to check that the performance of each compute node in the cluster performs as predicted. This is done by running the STREAM benchmarking program.

Perform the following on all compute nodes in the cluster:

  1. Set the compute nodes to run in a default frequency (The default CPU frequency can be set using the utility initialize_eas -f).
  2. Run STREAM on each compute node 10 times.
  3. Gather the performance value of the benchmark.

    The output of the STREAM benchmark is the triad value (the performance value).

  4. Calculate the average performance value of each compute node and compare it with the reference value.
Note: A node should not be used for energy aware scheduling if the measured performance is more than 4% lower than the reference value.
Note: The reference value is 70GB/s.

If a problem node is found after running the STREAM benchmarking program, you can:

  • Check that the firmware of the problem nodes is the same as other nodes.
  • Check that the threading mode (like Turbo or HT) is functioning on the problem nodes.
  • Check the current CPU frequency of the problem nodes.
  • Check the memory configuration of the problem nodes.

After performing the recommended checks, rerun the STREAM benchmark.