mpiexec scalability improved!

Garrick Staples garrick at usc.edu
Mon Apr 10 21:58:39 EDT 2006


On Mon, Apr 10, 2006 at 03:52:58PM -0700, Garrick Staples alleged:
> The initial processes are exec'd very quickly (a printf() before
> MPI_Init() gets back through mpiexec within a couple seconds), but the
> read_gm_startup_ports() loop can still stall for a few minutes.

I think this comes down to the fundamental difference between Myricom's
perl script and mpiexec: the perl script pre-forks to listen and reply
to remote processes before launching them; whereas mpiexec launches all
remote processes first and then starts accept()ing connections.

The perl script might be slower, but it succeeds 100% of the time.
mpiexec fails about 1/10 of the launches.

-- 
Garrick Staples, Linux/HPCC Administrator
University of Southern California
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://email.osc.edu/pipermail/mpiexec/attachments/20060410/fec238e8/attachment.bin


More information about the mpiexec mailing list