Can this patch be moved out of the ompi internals ? If possible it
should go in the MPI level functions somewhere. I looked into the mpi/
c/group_incl.c and we have a bunch of tests, but this particular one
is missing.
Long ago we decided to do most of the checking outside the internals
in order to have a clean [and fast] execution path once we get inside
the ompi functions. Of course, we suppose that all our internal calls
will never generate any problems (i.e. all the required checks will
be done prior to calling another function).
george.
On Aug 16, 2007, at 11:53 AM, Tim Prins wrote:
Mohamad,
2 process was plenty. Like I said, when running in debug mode, it
tends
to 'work' since memory is initialized to \0 and we fall through. In an
optimized build, looking at the mtt results it looks like it segfaults
about 10% of the time.
But if you apply the patch I sent, it will tell you when an invaild
lookup happens, which should be every time it runs.
Tim
Mohamad Chaarawi wrote:
Hey Tim,
I understand what you are talking about.
Im trying to reproduce the problem. How many processes are your
running
with, because with the default (4 for the group) it's passing..
Thanks,
Mohamad
On Thu, August 16, 2007 7:49 am, Tim Prins wrote:
Sorry, I pushed the wrong button and sent this before it was
ready....
Tim Prins wrote:
Hi folks,
I am running into a problem with the ibm test 'group'. I will
try to
explain what I think is going on, but I do not really understand
the
group code so please forgive me if it is wrong...
The test creates a group based on MPI_COMM_WORLD (group1), and a
group
that has half the procs in group1 (newgroup). Next, all the
processes
do:
MPI_Group_intersection(newgroup,group1,&group2)
ompi_group_intersection figures out what procs are needed for
group2,
then calls
ompi_group_incl, passing 'newgroup' and '&group2'
This then calls (since I am not using sparse groups)
ompi_group_incl_plist
However, ompi_group_plist assumes that the current process is a
member
of the passed group ('newgroup'). Thus when it calls
ompi_group_peer_lookup on 'newgroup', half of the processes get
garbage
back since they are not in 'newgroup'. In most cases, memory is
initialized to \0 and things fall through, but we get intermittent
segfaults in optimized builds.
Here is a patch to a error check which highlights the problem:
Index: group/group.h
===================================================================
--- group/group.h (revision 15869)
+++ group/group.h (working copy)
@@ -308,7 +308,7 @@
static inline struct ompi_proc_t* ompi_group_peer_lookup
(ompi_group_t
*group, int peer_id)
{
#if OMPI_ENABLE_DEBUG
- if (peer_id >= group->grp_proc_count) {
+ if (peer_id >= group->grp_proc_count || peer_id < 0) {
opal_output(0, "ompi_group_lookup_peer: invalid peer index
(%d)", peer_id);
Thanks,
Tim
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
