[gridengine users] Core Binding on Magny-Cours

[Dave Love]

"Esztermann, Ansgar" <Ansgar.Esztermann at mpi-bpc.mpg.de> writes:

> That's exactly what I'd expect...

Good.  I've only just encountered Magny-Cours, and I'm not sure I
understand them.

> The interface at /sys/devices/system/cpu/cpuN/topology/ doesn't know
> about NUMANodes, only about Sockets and cores. Thus, cores #0 and #4
> in the output above have the same core ID, and SGE interprets that as
> being one core with two threads.

Anyhow, sufficiently recent Linux does expose the information (under
/sys/devices/system/node/), it's just that PLPA doesn't use it by
design.

I guess that it would be a fairly straightforward project for someone to
try replacing PLPA with hwloc without needing to know much about GE
internals.  If anyone is interested, see
http://www.open-mpi.org/projects/hwloc/doc/v1.1.1/switchfromplpa.php
As well as helping with architectures like Magny-Cours, it would
introduce support for more platforms than Linux and Solaris.