On Jan 6, 2011, at 11:23 PM, Gilbert Grosdidier wrote:

lstopo

Machine (35GB)

NUMANode L#0 (P#0 18GB) + Socket L#0 + L3 L#0 (8192KB)

  L2 L#0 (256KB) + L1 L#0 (32KB) + Core L#0

    PU L#0 (P#0)

    PU L#1 (P#8)

  L2 L#1 (256KB) + L1 L#1 (32KB) + Core L#1

    PU L#2 (P#1)

    PU L#3 (P#9)

  L2 L#2 (256KB) + L1 L#2 (32KB) + Core L#2

    PU L#4 (P#2)

    PU L#5 (P#10)

  L2 L#3 (256KB) + L1 L#3 (32KB) + Core L#3

    PU L#6 (P#3)

    PU L#7 (P#11)

[snip]

Well, this might disprove my theory.  :-\  The OS indexing is not contiguous on the hyperthreads, so I might be wrong about what happened here.  Try this:

mpirun --mca mpi_paffinity_alone 1 hwloc-bind --get

You can even run that on just one node; let's see what you get.  This will tell us what each process is *actually* bound to.  hwloc-bind --get will report a bitmask of the P#'s from above.  So if we see 001, 010, 011, ...etc, then my theory of OMPI binding 1 proc per hyperthread (vs. 1 proc per core) is incorrect.

--
Jeff Squyres
jsquyres@cisco.com
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/