[mvapich-discuss] OpenFabrics
Di Domenico, Michael
mdidomenico at silverstorm.com
Wed Aug 30 22:56:58 EDT 2006
Kevin,
I saw similar weirdness running up to 4proc per machine. I've got Quad
Proc Itanium2's here. But I suspected it was MVAPICH to blame really,
seemd more of an OFED issue as I also tested with Scali and got weird
results as well.
Who makes the servers your running on? 2.6.12 kernel? What OS / Distro
are you running?
-----Original Message-----
From: Kevin Ball [mailto:kball at pathscale.com]
Sent: Wednesday, August 30, 2006 8:31 PM
To: Di Domenico, Michael
Cc: Sayantan Sur; mvapich-discuss at cse.ohio-state.edu
Subject: RE: [mvapich-discuss] OpenFabrics
On Wed, 2006-08-30 at 14:52, Di Domenico, Michael wrote:
> Sayantan,
>
> I don't recall seeing that error during my runs...
I get this error with both osu_bw and osu_bibw. Running some other,
related bw microbenchmarks I see this error if I'm running 1 process per
node, but with 2 or more processes per node I see segfaults instead, but
at around the same message size.
-Kevin
> Kevin,
>
> Can you describe the hardware your running on?
Itanium2 processors, MT25204 Mellanox cards with fw_ver 1.1.0. And as
mentioned before, OFED 1.1-rc2 with MVAPICH 0.9.7, on a 2.6.12 kernel.
-Kevin
>
> -----Original Message-----
> From: Sayantan Sur [mailto:surs at cse.ohio-state.edu]
> Sent: Wednesday, August 30, 2006 5:48 PM
> To: Kevin Ball
> Cc: Di Domenico, Michael; mvapich-discuss at cse.ohio-state.edu
> Subject: Re: [mvapich-discuss] OpenFabrics
>
> Kevin,
>
> >2097152 966.749701
> >0 - MPI_ISEND : Cannot free permanent data type
> >[0] [] Aborting Program!
> >mpirun_rsh: Abort signaled from [0]
> >1 - MPI_IRECV : Cannot free permanent data type
> >[1] [] Aborting Program!
> >done.
> >
> >
> That is weird. Are you using OSU bandwidth unmodified from our
website?
> As far as I recall, I had no problems with OSU benchmarks on IA64
> systems Michael provided access to. Michael, did you ever see this
> error?
>
> Thanks,
> Sayantan.
More information about the mvapich-discuss
mailing list