I am running into similar issues with both Mellanox and IBM HCAs.
On a node installed with RHEL6.2 and MLNX_OFED-1.5.3-3.0.0, there is a significant hit to locked memory when going with the device's max_cqe. Here, for comparison's sake is the memory utilization for a simple MPI process when using the new cq_size default, and when restricting it to 1500:
cq_size = max_cqe:
VmPeak: 348736 kB
VmSize: 348352 kB
VmLck: 292096 kB
VmHWM: 304896 kB
VmRSS: 304896 kB
VmData: 333504 kB
cq_size = 1500
VmPeak: 86720 kB
VmSize: 86336 kB
VmLck: 30080 kB
VmHWM: 42880 kB
VmRSS: 42880 kB
VmData: 71488 kB
For our Power systems using the IBM eHCA, the default value exhausts memory and we can't even run.
--Brad
I noticed this on a Solaris machine, I am not sure I have the same set up for Linux but I'll look and see if I can reproduce the same issue on Linux.
On 7/5/2012 5:47 PM, Shamis, Pavel wrote:I mentioned on the call that for Mellanox devices (+OFA verbs) this resource is really cheap. Do you run mellanox hca + OFA verbs ?(I'll reply because I know Terry is offline for the rest of the day) Yes, he does.I asked because SUN used to have own verbs driver.
--td
The heart of the question: is it incorrect to assume that we'll consume (num CQE * CQE size) registered memory for each QP opened?QP or CQ ? I think you don't want to assume anything there. Verbs (user and kernel) do their own magic there. I think Mellanox should address this question. Regards, Pasha _______________________________________________ devel mailing list devel@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/devel
--
Terry D. Dontje | Principal Software Engineer
Developer Tools Engineering | +1.781.442.2631
Oracle - Performance Technologies
95 Network Drive, Burlington, MA 01803
Email terry.dontje@oracle.com
_______________________________________________
devel mailing list
devel@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel