Re: Shared Memory usage in -machinefile
Reuti <[email protected]> Mon, 28 Mar 2005 02:21:25 +0200
| Newsgroups | gmane.network.myrinet.general |
|---|---|
| Message-ID | <[email protected]> |
Patrick, thx, I have guest account to a cluster with Myrinet and was wondering as = not=20 seeing anything with ipcs. Yes, I agree that IPC could be better implemen= ted to=20 be proper removed after a job. Somewhere I found some experimental code o= nce to=20 attach meta-information to this IPC stuff, to get it removed later. Cheers - Reuti Quoting Patrick Geoffray <[email protected]>: > Hi Reuti, >=20 > Reuti wrote: > > Question is: how is Myrinet behaving in this case? Different from the > default=20 > > MPICH, or could the above statement give under some circumstances a l= ess=20 > > optimal result? >=20 > node1:2 >=20 > will be expanded in >=20 > node1 > node1 >=20 > The shared memory support in MPICH-GM is very different from the=20 > -comm=3Dshared of MPICH-P4. When shared memory is enabled (the default)= =20 > all processes from the same MPI job running on the same host (hostname=20 > is the same) will use shared memory. We don't use IPC for allocated=20 > shared memory segments (IPC is a mess), so shared memory is either used= =20 > or not used, there is no other variation. >=20 > So, you can use ":n" or list the same node n times, there would be no=20 > difference regarding shared memory. >=20 > Patrick > -- >=20 > Patrick Geoffray > Myricom, Inc. > http://www.myri.com >=20