[mpich-discuss] Cannot build mpich2-1.0.8p1 (nemesis) with PGI8.0-4 on Linux x86_64

[Gus Correa]

Hi Rajeev, list

Rajeev Thakur wrote:
>>> We have dual-socket quad-core Opteron processor nodes (8 
>> cores/node).
>>> I am afraid ch3:sock may not be the best choice for this type of
>>> "fat" node, where shared memory shortcuts (memcpy ?)
>>> may work better than sockets.
> 
> You could try the ch3:ssm channel then.
> 
> Rajeev


Sure, and I did use ch3:ssm on a standalone multicore workstation
with good results.

However, for a Beowulf cluster, ch3:nemesis seems to promise the best
of all worlds: shared memory communication for intranode,
and a TCP/IP mechanism for internode.
In addition, a number of codes we run here require more than
8 processes, and will use more than one node at a time,
which cannot be done with ch3:ssm.

It is also a matter of convenience, otherwise I would have to
keep MPICH2 builds for ch3:ssm, ch3:sock, and ch3:nemesis,
which combined with different compilers (Gnu, Intel, PGI, and
hybrid compiler mixes) would give me too large a number of
libraries to build and maintain.
I would rather build MPICH2 with the most flexible
communication channel, which seems to be nemesis, right?

Also, I heard from the MPICH2 pros
(i.e. a gentleman called Rajeev Thakur and his team)
very convincing arguments to use ch3:nemesis.
And I believe them!  :)

Gus Correa
---------------------------------------------------------------------
Gustavo Correa
Lamont-Doherty Earth Observatory - Columbia University
Palisades, NY, 10964-8000 - USA
---------------------------------------------------------------------