Open MPI logo

Hardware Locality Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Hardware Locality Development mailing list

Subject: Re: [hwloc-devel] WHOLE_SYSTEM and cpusets (was: 1.3.1 and 1.4rc1)
From: Samuel Thibault (samuel.thibault_at_[hidden])
Date: 2011-12-14 11:29:18


Jeff Squyres, le Wed 14 Dec 2011 16:06:27 +0100, a écrit :
> - cpuset may not be equal to complete_cpuset

Yes, but not only in that case, see below.

> - cpuset is already the logical AND of allowed_cpuset and online_cpuset

Yes, that's explicitly enforced by the core

> - cpuset is comprised only of PUs that you can bind to

And thus this is right too.

> - (ditto for nodeset / complete_nodeset)

Ditto.

> And if you *do* supply WHOLE_SYSTEM, then
>
> - cpuset will be equal to complete_cpuset

Cf the doc: complete_cpuset also includes PUs for which we don't have
topology information.

> - cpuset is not necessarily the logical AND of allowed_cpuset and online_cpuset

Yes. As the documentation says, it comprises all PUs for which we have
topology information, be they allowed, offline, or not.

Samuel