5.26.10 - Resource Requests for Using Westmere Nodes On Pleiades

Over the weekend the effective computing capacity of Pleiades was increased by 44% with the addition of 24 racks of Westmere-based nodes. Despite the increase, jobs may still be waiting. A quick glimpse of the system loading (using /u/scicon/tools/bin/node_stats.sh) shows why:

%/u/scicon/tools/bin/node_stats.sh
						
Pleiades Nodes Total: 9213
Pleiades Used Total: 7786

Harpertown Total: 5887 Used: 5829 Free: 58
Nehalem    Total: 1279 Used: 1262 Free: 17
Westmere   Total: 2047 Used: 695 Free: 1352

%

In this snapshot, the Harpertown and Nehalem nodes are essentially at capacity. However, there is room for jobs to run on Westmere nodes.

If you have been using Harpertowns and want to try Westmeres, please note the following:

• Westmere nodes have 3x the memory of Harpertown nodes
• Westmere nodes have 12 cores instead of 8

In light of this, you should adjust your resource request. While there are many possibilities to explore, we recommend starting with the following. If your current request is:

#PBS -l select=M:ncpus=P:model=har

then try:

#PBS -l select=N:ncpus=Q:model=wes

where N = 2*M/3 rounded up and Q = M*P/N rounded up.

If you have been running with ncpus=8 on Harpertown, you could also try "hyperthreading" where the execution will have more than one process per core. In this case, you can change the value of N in the above example to N = M/3 rounded up; the value of Q is computed as before: Q = M*P/N.

For more information about the different node architectures, please review detils on Pleiades hardware.

If you have any questions or problems, please send email to support@nas.nasa.gov or call 1-800-331-8737 or 1-650-604-4444.