Open MPI logo

Hardware Locality Development Mailing List Archives

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

Subject: Re: [hwloc-devel] [hwloc-svn] svn:hwloc r1255
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2009-10-29 08:32:09

On Oct 29, 2009, at 7:38 AM, Samuel Thibault wrote:

> I'm just wondering: maybe by default hwloc should try strict binding
> and if it fails, if the strict flag is set, error out, else revert to
> non-strict binding. That way only people who really want absolute
> strict
> binding and fail if it is not available will use the hwloc strict
> flag.

I think there's 3 cases:

1. If I specify STRICT, then I would expect strict binding to occur,
or the function fail if the OS can't deliver it. I think we agree on
this case.

2. If I specify LOOSE (or whatever the opposite of STRICT is -- and I
see there's currently no flag for that), then I think you get whatever
the OS gives you -- even if it happens to be strict. I.e., strict is
a subset of loose. We might agree on this case, too...?

3. If I don't specify anything, I think this is an open question as to
what happens here. I'd be in favor of making LOOSE be the default.
What I'm not in favor of is having it try STRICT and if that fails,
try LOOSE. It feels to me like if the user wants STRICT, they should
try it. And if it fails, they should be notified of that (through the
function failing) and then they can choose how to react.

Jeff Squyres