Open MPI logo

Hardware Locality Development Mailing List Archives

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

Subject: Re: [hwloc-devel] multinode/network topology status
From: Samuel Thibault (samuel.thibault_at_[hidden])
Date: 2011-10-04 10:51:34


Hello,

Brice Goglin, le Wed 21 Sep 2011 08:54:52 +0200, a écrit :
> Example of C interface use (assembling 3 nodes in a global topology with
> 3 levels of switches):
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/tests/hwloc_custom.c#L23
>
> C interface doc:
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/include/hwloc.h#L1904
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/include/hwloc.h#L803
>
> Manpages of the command-line tools (with examples at the end):
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/utils/hwloc-assembler.1in
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/utils/hwloc-assembler-remote.1in
>
> General doc about all this:
> https://svn.open-mpi.org/trac/hwloc/browser/branches/custom/doc/hwloc.doxy#L1224

This looks good. hwloc-assembler-remote should probably have a
--lstopo-path option to specify where to find lstopo, since the host
from which we run it may not have the same hwloc installation as the
target nodes. I'm actually unsure whether it really should default to
$bindir/lstopo and not simply to lstopo.

There is an issue with the distance code that appears when there are
several machines with NUMA distances: the distance code assumes that the
root of the topology has cpusets to be compared with, and quickly
crashes on insert_by_cpuset or get_obj_covering_cpuset calls.

Samuel