February 2010 - Mpich2-intel - mailman3.cels.anl.gov

Pmi_proxy issue
by Kiryanov, Yury 13 Apr '10

13 Apr '10

Hello everybody, Pavan, One issue which I observed with mpich2-1.3a1. It is related to launcing of pmi_proxy processes. I marked bold the unexpected behaviour. 1. I use the following hostsfile: host0 host1 host0 host1 My command line is: > mpiexec --print-rank-map -f hostsfile -n 4 ./test (host0:0) (host1:1) (host0:2) (host1:3) Hello world: rank 0 of 4 running on host0 Hello world: rank 1 of 4 running on host1 Hello world: rank 2 of 4 running on host0 Hello world: rank 3 of 4 running on host1 I looked into host0 and host1 and observed two pmi_proxy processes on each node. I guess it should be one pmi_proxy process on each node. Is it a bug or expected behavior? Other scenarios which I use worked as expected: 2. hostsfile: host0 host0 host1 host1 > mpiexec --print-rank-map -f hostsfile -n 4 ./test (host0:0,1) (host1:2,3) Hello world: rank 0 of 4 running on host0 Hello world: rank 1 of 4 running on host0 Hello world: rank 2 of 4 running on host1 Hello world: rank 3 of 4 running on host1 3. hostsfile: host0 host1 > mpiexec --print-rank-map -f hostsfile -n 4 ./test (host0:0,2) (host1:1,3) Hello world: rank 0 of 4 running on host0 Hello world: rank 1 of 4 running on host1 Hello world: rank 2 of 4 running on host0 Hello world: rank 3 of 4 running on host1 Thank you. -- Best regards, Yury Kiryanov phone: 8-83130-70700 (ext. 2531) i-net : 8313-2531 e-mail: yury.kiryanov(a)intel.com -------------------------------------------------------------------- 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 3

DOUBLE PRECISION
by Kiryanov, Yury 12 Feb '10

12 Feb '10

Hello, Could you explain why configure substitutes the 'DOUBLE PRECISION' to 'REAL*8' declaration of MPI_WTIME, MPI_WTICK into mpif.h and mpi_base.f90? It has been changed into mpich2-1.0.8. I'm asking about it because the MPI-2 says that it should be declared as DOUBLE PRECISION. Thank you. -- Best regards, Yury Kiryanov phone: 8-83130-70700 (ext. 2531) i-net : 8313-2531 e-mail: yury.kiryanov(a)intel.com -------------------------------------------------------------------- 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 1

Re: [mpich2-intel] Hydra testing
by Kiryanov, Yury 10 Feb '10

10 Feb '10

Dear Pavan, I've tested a little the latest version of hydra (rev.6184). Below are results. 1. As I see hydra supports -iface <net_interface>. One question related this option. I noticed there are no difference in startup time (MPI_Init) by default through ethernet in comparison with ib interface. I rerun it with -verbose option and saw that ethernet uses for both variants of PMI connections. Are you going to change it or to add separate option to use interface specified by -iface for PMI connections? I guess it would be great to have this functionality to improve startup time on big systems. 2. I didn't find proxy options -boot-proxies, -boot-foreground-proxies, -shutdown-proxies, -use-persistent. Did you drop them? 3. I tried to use -binding option and it hanged (output below). I interrupted it by Ctrl+c. When I use -demux select with -binding option I could not interrupt it by Ctrl+C. mpiexec.hydra -f $PBS_NODEFILE -binding rr -n 8 ./test [mpiexec@en138] HYDT_dmxu_poll_wait_for_event (./tools/demux/demux_poll.c:71): assert (!(pollfds[i].revents & ~POLLIN & ~POLLOUT & ~POLLHUP)) failed [mpiexec@en138] HYD_pmci_wait_for_completion (./pm/pmiserv/pmiserv_pmci.c:142): error waiting for demux event [mpiexec@en138] main (./ui/mpich/mpiexec.c:230): process manager error waiting for completion 4. I've tested -rmk pbs option and it works correct. 5. -bootstrap option works correct, but there are some problems with -bootstrap-exec mpiexec.hydra -rmk pbs -bootstrap-exec rsh -n 8 ./test [proxy@en138] match_arg (./utils/args/args.c:122): unrecognized argument bootstrap-exec [proxy@en138] HYDU_parse_array (./utils/args/args.c:140): argument matching returned error [proxy@en138] HYD_pmcd_pmip_get_params (./pm/pmiserv/pmip_utils.c:423): error parsing input array [proxy@en138] main (./pm/pmiserv/pmip.c:146): bad parameters passed to the proxy [proxy@en150] match_arg (./utils/args/args.c:122): unrecognized argument bootstrap-exec [proxy@en150] HYDU_parse_array (./utils/args/args.c:140): argument matching returned error [proxy@en150] HYD_pmcd_pmip_get_params (./pm/pmiserv/pmip_utils.c:423): error parsing input array [proxy@en150] main (./pm/pmiserv/pmip.c:146): bad parameters passed to the proxy EXIT STRING: Hangup (signal 1) Thank you. -- Yury -----Original Message----- From: Yalozo, Artem V Sent: Tuesday, February 02, 2010 9:58 AM To: Pavan Balaji Cc: Kiryanov, Yury Subject: RE: Hydra testing Hi Pavan, Thank you very much. We will try out the latest tarball and let you know of any issues found. Regards, Artem -----Original Message----- From: Pavan Balaji [mailto:[email protected]] Sent: Tuesday, February 02, 2010 2:37 AM To: Yalozo, Artem V Subject: Hydra testing Hi Artem, We are just about to release mpich2-1.3a1 which has Hydra as the default process manager. For this, we want to make sure that all the features requested by Intel are supported in it (including tunneling PMI connections through proxies, etc.). Would you be willing to try out the nightly snapshots of Hydra from here (http://www.mcs.anl.gov/research/projects/mpich2/downloads/tarballs/nightly/…) to make sure everything is OK? We would certainly like to hear about any bugs or places it doesn't work, but we would also be interested in issues that are not really bugs but just annoying to have. Thanks, -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji -------------------------------------------------------------------- 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 6