Posted by Serguei Patchkovskii on June 02, 2009 at 00:14:58:
In Reply to: Re: Weird compilation problems with GA in MOLCAS 7.2 posted by Valera Veryazov on June 01, 2009 at 23:53:35:
: Thanks for a good work in finding proper installation flags for GA, but now I have a question.. Can you
: compare the benchmarks: your tuned version vs. 'vanilla' TCGMSG? Molcas currently uses so rare
: interactions between nodes...
Unfortunately, I wiped my tcgmsg installation after I got mpi-2 setup running - so can't give the
exact numbers.
One change I noticed is that for multi-node jobs the wall-clock times are closer to cpu times,
due to decreased startup time. OpenMPI has an accelerated start-up facility through the orted
daemon, while tcgmsg has to go through ssh authentication cycle each time a process starts.
For a medium-size cluster using NFS-mounted system partitions, this gives a (small) improvement
in wall-clock times.
Of course, the gain is significant only for very small jobs - which arguably should be run on a single
node in the first place.