Parallel HDF5: "make check" hangs when running t_mpi - parallel-processing

Question

I've being struggling to get the parallel HDF5 to work on the cluster for a whole week but without any progress. I wish if anyone could help me with this. Thanks!
I'm building the Parallel HDF5 (hdf5-1.8.15-patch1) on a lustre file system with RedHat Enterprise Linux 5.5 x86_64 OS. I tried to compile it with both impi 4.0.2 and openmpi 1.8 and it succeeded without any errors. When I "make check", both of them passed the serial tests, but hung immediately after entering the parallel tests (t_mpi, in particular). Eventually, I had to ctrl+C to end it. Here is the output:
lijm#c01b03:~/yuan/hdf5-1.8.15-patch1/testpar$ make check
CC t_mpi.o
t_mpi.c: In function ‘test_mpio_gb_file’:
t_mpi.c:284: warning: passing argument 1 of ‘malloc’ with different width due to prototype
t_mpi.c:284: warning: request for implicit conversion from ‘void *’ to ‘char *’ not permitted in C++
t_mpi.c: In function ‘test_mpio_1wMr’:
t_mpi.c:465: warning: passing argument 2 of ‘gethostname’ with different width due to prototype
t_mpi.c: In function ‘test_mpio_derived_dtype’:
t_mpi.c:682: warning: declaration of ‘nerrors’ shadows a global declaration
t_mpi.c:37: warning: shadowed declaration is here
t_mpi.c:771: warning: passing argument 5 of ‘MPI_File_set_view’ discards qualifiers from pointer target type
t_mpi.c:798: warning: passing argument 2 of ‘MPI_File_set_view’ with different width due to prototype
t_mpi.c:798: warning: passing argument 5 of ‘MPI_File_set_view’ discards qualifiers from pointer target type
t_mpi.c:685: warning: unused variable ‘etypenew’
t_mpi.c:682: warning: unused variable ‘nerrors’
t_mpi.c: In function ‘main’:
t_mpi.c:1104: warning: too many arguments for format
t_mpi.c: In function ‘test_mpio_special_collective’:
t_mpi.c:991: warning: will never be executed
t_mpi.c:992: warning: will never be executed
t_mpi.c:995: warning: will never be executed
t_mpi.c: In function ‘test_mpio_gb_file’:
t_mpi.c:229: warning: will never be executed
t_mpi.c:232: warning: will never be executed
t_mpi.c:237: warning: will never be executed
t_mpi.c:238: warning: will never be executed
t_mpi.c:253: warning: will never be executed
t_mpi.c:258: warning: will never be executed
t_mpi.c:259: warning: will never be executed
t_mpi.c:281: warning: will never be executed
t_mpi.c:246: warning: will never be executed
t_mpi.c:267: warning: will never be executed
t_mpi.c:319: warning: will never be executed
t_mpi.c:343: warning: will never be executed
t_mpi.c:385: warning: will never be executed
t_mpi.c:389: warning: will never be executed
t_mpi.c:248: warning: will never be executed
t_mpi.c:269: warning: will never be executed
t_mpi.c: In function ‘main’:
t_mpi.c:1143: warning: will never be executed
t_mpi.c:88: warning: will never be executed
t_mpi.c:102: warning: will never be executed
t_mpi.c:133: warning: will never be executed
t_mpi.c:142: warning: will never be executed
CCLD t_mpi
make t_mpi testphdf5 t_cache t_pflush1 t_pflush2 t_pshutdown t_prestart t_shapesame
make[1]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[1]: `t_mpi' is up to date.
make[1]: `testphdf5' is up to date.
make[1]: `t_cache' is up to date.
make[1]: `t_pflush1' is up to date.
make[1]: `t_pflush2' is up to date.
make[1]: `t_pshutdown' is up to date.
make[1]: `t_prestart' is up to date.
make[1]: `t_shapesame' is up to date.
make[1]: Leaving directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make check-TESTS
make[1]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[2]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[3]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[3]: Nothing to be done for `_exec_check-s'.
make[3]: Leaving directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[2]: Leaving directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
make[2]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
===Parallel tests in testpar begin Thu Jun 11 22:07:48 CST 2015===
**** Hint ****
Parallel test files reside in the current directory by default.
Set HDF5_PARAPREFIX to use another directory. E.g.,
HDF5_PARAPREFIX=/PFS/user/me
export HDF5_PARAPREFIX
make check
**** end of Hint ****
make[3]: Entering directory `/home/lijm/yuan/hdf5-1.8.15-patch1/testpar'
============================
Testing t_mpi
============================
t_mpi Test Log
============================
===================================
MPI functionality tests
===================================
Proc 1: hostname=c01b03
Proc 2: hostname=c01b03
Proc 3: hostname=c01b03
Proc 5: hostname=c01b03
--------------------------------
Proc 0: *** MPIO 1 write Many read test...
--------------------------------
Proc 0: hostname=c01b03
Proc 4: hostname=c01b03
Command exited with non-zero status 255
0.08user 0.01system 0:37.65elapsed 0%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+5987minor)pagefaults 0swaps
make[3]: *** [t_mpi.chkexe_] Error 1
make[2]: *** [build-check-p] Interrupt
make[1]: *** [test] Interrupt
make: *** [check-am] Interrupt
The above outputs of two MPI implementations are the same, but openmpi also outputs the warning:
WARNING: It appears that your OpenFabrics subsystem is configured to only allow registering part of your physical memory. This can cause MPI jobs to run with erratic performance, hang, and/or crash.
I've searched for this problem. But I don't think it could be the cause for the hanging, the reason is stated at the end.
I've tried to locate the place where it hangs. What I found is that it always gets stuck on the first collective function it meets. For example, in t_mpi. it firstly hangs at:
MPI_File_delete(filename, MPI_INFO_NULL); (line 477),
in test_mpio_1wMr. And if I comment out this line, it gets stuck at MPI_File_open just below. But I'm not sure about what happened inside these functions.
There is another thing I noticed. The folder of HDF5 where I do the "make" is in a NFS file system, and I can only access the lustre through a particular folder located somewhere else. So, I found that the test runs pretty well if I don't set the HDF5_PARAPERFIX to my lustre folder, since the test is performed locally by default. So, I suppose it should be a issue related with the lustre itself, not the limit of memory?
Thank you!

Answer 1

It's hard to say what's going on here.
It may be that you are applying "generic unix file system" to your lustre driver. Intel MPI requires two environment variables (I_MPI_EXTRA_FILESYSTEM and I_MPI_EXTRA_FILESYSTEM_LIST) to use lustre-optimized code paths: (see https://press3.mcs.anl.gov/romio/2014/06/12/romio-and-intel-mpi/ for more details).
You'll have to explicitly request lustre support when you build OpenMPI, too.
It would help a lot if you can attach a debugger to one or more of the stuck processes to see where its hanging. stuck on an i/o routine? stuck in communication?