Experiment with Processor Topology/Affinity/Pinning

[haesbaert]

I've recently released https://github.com/haesbaert/ocaml-processor which exposes the processor topology and allows for pinning of processes/domains to a set of cpus.

There were some discussions in Paris concerning which tests are affected more/less by the presence of hyper-threading and/or sockets and other NUMA artifacts.

Using ocaml-processor might expose some performance properties of different benchmarks when the affinity is changed, as in:
"What happens with test X when I restrict socket count, or isolate hyper-threading".

I'm obviously available for any help needed with the setup.

ccing @Sudha247 who suggested opening an issue.

[shakthimaan]

At present, the .bench file produced by Sandmark has a header entry with kernel, version and CPU architecture information. The ocaml-processor library will be useful to augment the meta-data. Thanks for filing the enhancement request.

[haesbaert]

At present, the .bench file produced by Sandmark has a header entry with kernel, version and CPU architecture information. The ocaml-processor library will be useful to augment the meta-data. Thanks for filing the enhancement request.

Awesome !

I think it could also be useful to see how dual socket machines perform if you restrict sockets, or what happens with tests if you don't use hyper-threading as well