Thread binding is wrong when multiple ranks per node are used.

[Raffaele Solcà]

When using multiple ranks per node the threads of the different ranks are bound to the same cores. Is there a way to modify this behaviour?

[George Bosilca]

Yes, but it is not very easy. One can use the MCA parameter parsec_bind to describe the binding for each process, but this will force you to write the binding by hand. If you set this MCA parameter to file:my_bindings, then each process will load these bindings find it's rank inside and follow the bindings as described in there (with the format start:stop:step).

This has been on my radar for a while. I'll be at SC next week, but I'll try to resurrect a patch I have lingering around.