Feel free to do so.
Thanks,
george.
On Oct 30, 2009, at 16:46 , Eugene Loh wrote:
Thanks. So, you'll put back the change?
George Bosilca wrote:
The progress functions do not need to return an error code. If
there is an error they should propagate it back through the
descriptors. The only meaning of the return code of the progress
functions is to know if any event had happened during this round
of progress. The opal_output use it to trigger the yield (if one
if necessary).
Anyway, this is a good catch. We're doing a big mixup there.
Attached you will find a patch that clean this problem. As
expected, there is no performance impact ...
Here is the patch:
Index: btl_sm_component.c
===================================================================
--- btl_sm_component.c (revision 22176)
+++ btl_sm_component.c (working copy)
@@ -361,7 +361,7 @@
sm_fifo_t *fifo = NULL;
mca_btl_sm_hdr_t *hdr;
int my_smp_rank = mca_btl_sm_component.my_smp_rank;
- int peer_smp_rank, j, rc = 0;
+ int peer_smp_rank, j, rc = 0, events = 0;
/* first, deal with any pending sends */
/* This check should be fast since we only need to check one
variable. */
@@ -399,7 +399,7 @@
continue;
}
- rc++;
+ events++;
/* dispatch fragment by type */
switch(((uintptr_t)hdr) & MCA_BTL_SM_FRAG_TYPE_MASK) {
case MCA_BTL_SM_FRAG_SEND:
@@ -480,5 +480,5 @@
break;
}
}
- return rc;
+ return events;
}
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
