On Wed, Jan 17, 2007 at 08:55:31AM -0700, Brian W. Barrett wrote: >On Jan 17, 2007, at 2:39 AM, Gleb Natapov wrote: >> On Wed, Jan 17, 2007 at 04:12:10AM -0500, Robin Humble wrote: >>> basically I'm seeing wildly different bandwidths over InfiniBand 4x DDR >>> when I use different kernels. >> Try to load ib_mthca with tune_pci=1 option on those kernels that are >> slow. >when an application has high buffer reuse (like NetPIPE), which can >be enabled by adding "-mca mpi_leave_pinned 1" to the mpirun command >line.
thanks! :-) tune_pci=1 makes a huge difference at the top end, and -mca mpi_leave_pinned 1 adds lots of midrange bandwidth. latencies (~4us) and the low end performance are all unchanged. see attached for details. most curves are for 2.6.19.2 except the last couple (tagged as old) which are for 2.6.9-42.0.3.ELsmp and for which tune_pci changes nothing. why isn't tune_pci=1 the default I wonder? files in /sys/module/ib_mthca/ tell me it's off by default in 2.6.9-42.0.3.ELsmp, but the results imply that it's on... maybe PCIe handling is very different in that kernel. is ~10Gbit the best I can expect from 4x DDR IB with MPI? some docs @HP suggest up to 16Gbit (data rate) should be possible, and I've heard that 13 or 14 has been achieved before. but those might be verbs numbers, or maybe horsepower >> 4 cores of 2.66GHz core2 is required? >It would be interesting to know if the bandwidth differences appear >when the leave pinned protocol is used. My guess is that they will yeah, it definitely makes a difference in the 10kB to 10mB range. at around 100kB there's 2x the bandwidth when using pinned. thanks again! > Brian Barrett > Open MPI Team, CCS-1 > Los Alamos National Laboratory how's OpenMPI on Cell? :) cheers, robin