March 2011 - Mpich2-intel - mailman3.cels.anl.gov

DRMAA integration into Hydra PM
by Kiryanov, Yury 23 Mar '11

23 Mar '11

Hello everybody, The question is about job schedulers integration. As I heard DRMAA provides API interface compatible with widely used job schedulers (such as PBS, LSF, SGE, LoadLeveler). I am not an expert of DRMAA, just read some articles and I haven't tried to use it yet. Have you examined or were you thinking about a possibility to be integrated with DRMAA API interface into Hydra? The link to reference is http://en.wikipedia.org/wiki/DRMAA Thank you. -- Best regards, Yury Kiryanov phone: 8-83130-70700 (ext. 2531) i-net : 8313-2531 e-mail: yury.kiryanov(a)intel.com<mailto:[email protected]> -------------------------------------------------------------------- Closed Joint Stock Company Intel A/O Registered legal address: Krylatsky Hills Business Park, 17 Krylatskaya Str., Bldg 4, Moscow 121614, Russian Federation This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.

2 4

Scalability issue under Slurm
by Kiryanov, Yury 05 Mar '11

05 Mar '11

Hello Pavan, We've observed an issue under Slurm on big cluster with Hydra related to scalability. Seg fault occurs due to memory corruption (overrun memory bound). The issue is in unwind hosts list function which converts the $SLURM_NODELIST env. var. to ordinary host list (slurm_query_node_list.c: group_to_individaul_nodes, HYDT_bscd_slurm_query_node_list functions). HYD_NUM_TMP_STRINGS is used there for temporary storages, and sometimes it is not enough, for example if SLURM_NODELIST contains > 1000 hosts, for example "host[00001-10000]" or "host0001,host0002,...,host1030" or something like that. -- Best regards, Yury Kiryanov phone: 8-83130-70700 (ext. 2531) i-net : 8313-2531 e-mail: yury.kiryanov(a)intel.com<mailto:[email protected]> -------------------------------------------------------------------- Closed Joint Stock Company Intel A/O Registered legal address: Krylatsky Hills Business Park, 17 Krylatskaya Str., Bldg 4, Moscow 121614, Russian Federation This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.

2 9

mpi_lookup_name_ fortran wrapper issue
by Kiryanov, Yury 01 Mar '11

01 Mar '11

Hello everybody, We have run into an issue on mpi/f77/spawn/connaccf.f, mpi/f90/spawn/connaccf.f90 tests. The issue was observed on SLES 11 with Gfortran 4.3.2 compiler. These tests failed with message: Wrong size for intercomm = 2 The root cause of this issue is in mpi_lookup_name_ wrapper call. P3 variable is not initialized into wrapper function and sometimes memory is corrupted by this wrapper function if MPI_Lookup_name completes unsuccessfully. The memory corruption occurs if the allocated memory for p3 doesn't contain '\0'. The initialization of *p3 ='\0' solves this issue. -- Best regards, Yury Kiryanov phone: 8-83130-70700 (ext. 2531) i-net : 8313-2531 e-mail: yury.kiryanov(a)intel.com<mailto:[email protected]> -------------------------------------------------------------------- Closed Joint Stock Company Intel A/O Registered legal address: Krylatsky Hills Business Park, 17 Krylatskaya Str., Bldg 4, Moscow 121614, Russian Federation This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.

3 5