On 02/05/13 13:20, Eugene Loh wrote:
On 02/05/13 00:30, Siegmar Gross wrote:
now I can use all our machines once more. I have a problem on
Solaris 10 x86_64, because the mapping of processes doesn't
correspond to the rankfile.
A few comments.
First of all, the heterogeneous environment had nothing to do with this
(as you have just confirmed). You can reproduce the problem so:
% cat myrankfile
rank 0=mynode slot=0:1
% mpirun --report-bindings --rankfile myrankfile hostname
[mynode:5150] MCW rank 0 bound to socket 0[core 0-3]: [B B B B] (slot
list 0:1)
Anyhow, that's water under the bridge at this point.
Next, and you might already know this, you can't bind arbitrarily on
Solaris. You have to bind to a locality group (lgroup) or an individual
core. Sorry if that's repeating something you already knew. Anyhow,
your problem cases are when binding to a single core. So, you're all
right (and OMPI isn't).
Finally, you can check the actual binding so:
% cat check.c
#include <sys/types.h>
#include <sys/processor.h>
#include <sys/procset.h>
#include <stdio.h>
int main(int argc, char **argv) {
processorid_t obind;
if ( processor_bind(P_PID, P_MYID, PBIND_QUERY, &obind) != 0 ) {
printf("ERROR\n");
} else {
if ( obind == PBIND_NONE ) printf("unbound\n");
else printf("bind to %d\n", obind);
}
return 0;
}
% cc check.c
% mpirun --report-bindings --rankfile myrankfile ./a.out
I can reproduce your problem on my Solaris 11 machine (rankfile
specifies a particular core but --report-bindings shows binding to
entire node), but the test problem shows binding to the core I specified.
So, the problem is in --report-bindings? I'll poke around some.
I'm thinking the issue is in openmpi-1.6.3-debug/opal/mca/hwloc/hwloc132/hwloc/src/topology-solaris.c . First, we call
hwloc_solaris_set_sth_cpubind() to perform the binding. Again, there are two mechanisms for doing so: lgroup/lgrp for an entire
locality group or processor_bind to a specific core. In our case, we use the latter. Later, we call
hwloc_solaris_get_sth_cpubind() to check what binding we have. We check lgroups, but we don't check for processor_binding.
Sorry for the dumb question, but who maintains this code? OMPI, or upstream in
the hwloc project? Where should the fix be made?
