[mvapich-discuss] Segfault when running osu benchmarks on PSM2 hardware

[Brennan, Sean M]

Hi,
I get a SIGSGEV at psm_queue.c line 125 using slurm/pmi2 on
omnipath/psm2, only for PGI compiled code.

My build:
module load pgi/15.10 (or pgi/16.4)

./configure --prefix=/opt/mvapich2/2.2_pgi-16.4 --enable-f77 --enable-
fc --enable-cxx --enable-shared --enable-sharedlibs=gcc --enable-g=dbg
--enable-debuginfo --with-mpe --with-ch3-rank-bits=32 --with-
device=ch3:psm --with-psm2=/usr --enable-romio --with-file-
system=ufs+nfs+lustre --with-pmi=pmi2 --with-pm=slurm

make  && make install

To run in an interactive session:
srun --mpi=pmi2 -n 2 /opt/mvapich2/2.2_pgi-15.10/libexec/osu-micro-
benchmarks/mpi/pt2pt/osu_latency

Fails on all benchmarks.

My best guess is that PGI does not initialize the variables in
psm_progress_wait() properly/as expected; gblstatus.context is set to
0x1. (debugging with DDT)

Multiple versions of the gcc and intel compilers have no problems with
the same build invocation.

Suggestions?
Thanks.

-- 
############################

Sean M. Brennan, Ph.D.

High Performance Computing Systems
Los Alamos National Laboratory
PO Box 1663, MS T080
Los Alamos, NM 87545

(505) 667-1092, fax: (505) 667-7665