More information.
--mca mpi_paffinity_alone 0
still fails:
...
OMPI_MCA_mpi_paffinity_alone=0
OMPI_MCA_mpi_paffinity_processor=3
...
and it gets bound.
May I ask I question about the code, which I scanned.
I see in ompi/runtime/ompi_mpi_params.c you do a:
mca_base_param_reg_int_name("mpi", "paffinity_alone",
"If nonzero, assume that this job is the only
(set of) process(es) r
unning on each node and bind processes to processors, starting with processor ID
0",
false, false,
(int) ompi_mpi_paffinity_alone, &value);
ompi_mpi_paffinity_alone = OPAL_INT_TO_BOOL(value);
and in opal/mca/base/mca_base_param.c
/*
* Register an integer MCA parameter that is not associated with a
* component
*/
int mca_base_param_reg_int_name(const char *type,
const char *param_name,
const char *help_msg,
bool internal,
bool read_only,
int default_value,
int *current_value)
{
int ret;
mca_base_param_storage_t storage;
mca_base_param_storage_t lookup;
storage.intval = default_value;
ret = param_register(type, NULL, param_name, help_msg,
MCA_BASE_PARAM_TYPE_INT, internal, read_only,
&storage, NULL, NULL, &lookup);
if (ret >= 0 && NULL != current_value) {
*current_value = lookup.intval;
}
return ret;
}
So, it appears that the "current_value" returned from lookup.intval may not
happen
and you just get the "ret" value.
In the mca_base_param_reg_int_name call in ompi_mpi_params.c you don't check the
return code, so there may be junk in value? However, I don't see that explicitly
setting --mca mpi_paffinity_alone 0 would fail?
Regards,
Mostyn
P.S. I hope this doesn't seem too presumptious.
On Sun, 22 Apr 2007, Mostyn Lewis wrote:
Using a lateish SVN, 1.3a1r14155 (circa April 1st), on a SuSE SLES10
opteron system which has 2 dual core opterons per node, I can't seem
to disable processor affinity? I have a small test program which call
something else to detect CPU binding and whatever I've done it's still
bound?
tried:
--mca mpi_paffinity_processor -1
--mca mpi_paffinity_processor -99
--mca mpi_paffinity_processor fred (desperation - still bound!)
Any help?
Regards,
Mostyn
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
