[Beowulf] Optimizations for smaller scale linux clusters

[Trevor Gale]

Hello all,

I am building a small scale cluster (most likely around 8-10 nodes). I was wondering if anyone has any general optimizations that they like to implement on their clusters. We are using CentOS 6.6, with an IB network. Right now I have a test cluster (2 nodes) that we are using to test any optimizations.

I’ve almost finished reading through the performance tuning guide for red hat but I figured I’d ask here to see if anyone has any suggestions/comments.

Right now I’m thinking of doing the following.

-Configuring HugeTLB to reduce TLB misses
-Configuring overcommit_memory to increase performance for memory intensive tasks
-Configuring high “swappiness” so that inactive processes are aggressively swapped out of memory to make room for applications

We are also thinking about setting up a ethernet network so that all control can be run over it to leave the IB for applications exclusively, and limiting linux to running on one processor to free up all other resources for application runs.

Thanks,
Trevor