On Sep 16, 2008, at 11:18 PM, Matthias Hovestadt wrote:
Hi!
Since I am interested in fault tolerance, checkpointing and
restart of OMPI is an intersting feature for me. So I installed
BLCR 0.7.3 as well as OMPI from SVN (rev. 19553). For OMPI
I followed the instructions in the "Fault Tolerance Guide"
in the OMPI wiki:
./autogen.sh
./configure --with-ft=cr --enable-ft-thread --enable-mpi-threads
make -s
This gave me an OMPI version with checkpointing support, so I
started testing. The good news is: I am able to checkpoint and
restart applications. The bad news is: checkpointing a restarted
application fails.
In detail:
1) Starting the application
ccs@grid-demo-1:~$ ompi-clean
ccs@grid-demo-1:~$ mpirun -np 2 -am ft-enable-cr yafaray-xml
yafaray.xml
This starts my MPI-enabled application without any problems.
2) Checkpointing the application
First I queried the PID of the mpirun process:
ccs@grid-demo-1:~$ ps auxww | grep mpirun
ccs 13897 0.4 0.2 63992 2704 pts/0 S+ 04:59 0:00
mpirun -np 2 -am ft-enable-cr yafaray-xml yafaray.xml
Then I checkpointed the job, terminating it directly:
ccs@grid-demo-1:~$ ompi-checkpoint --term 13897
Snapshot Ref.: 0 ompi_global_snapshot_13897.ckpt
ccs@grid-demo-1:~$
The application indeed terminated:
----------------------------------------------------------------------
----
mpirun noticed that process rank 0 with PID 13898 on node grid-
demo-1.cit.tu-berlin.de exited on signal 0 (Unknown signal 0).
----------------------------------------------------------------------
----
2 total processes killed (some possibly by mpirun during cleanup)
The checkpoint command generated a checkpoint dataset
of 367MB size:
ccs@grid-demo-1:~$ du -s -h ompi_global_snapshot_13897.ckpt/
367M ompi_global_snapshot_13897.ckpt/
ccs@grid-demo-1:~$
3) Restarting the application
For restarting the application, I first executed ompi-clean,
then restarting the job with preloading all files:
ccs@grid-demo-1:~$ ompi-clean
ccs@grid-demo-1:~$ ompi-restart --preload
ompi_global_snapshot_13897.ckpt/
Restarting works pretty fine. The jobs restarts from the
checkpointed state and continues to execute. If not interrupted,
it continues until its end, returning a correct result.
However, I observed one weird thing: restarting the application
seemed to have the checkpoint dataset changed. Moreover, two new
directories have been created at restart time:
4 drwx------ 3 ccs ccs 4096 Sep 17 05:09
ompi_global_snapshot_13897.ckpt
4 drwx------ 2 ccs ccs 4096 Sep 17 05:09 opal_snapshot_0.ckpt
4 drwx------ 2 ccs ccs 4096 Sep 17 05:09 opal_snapshot_1.ckpt
The ('opal_snapshot_*.ckpt') directories are an artifact of the --
preload option. This option will copy the individual checkpoint to
the remote machine before executing.
4) Checkpointing again
Again I first looked for the PID of the running mpirun process:
ccs@grid-demo-1:~$ ps auxww | grep mpirun
ccs 14005 0.0 0.2 63992 2736 pts/1 S+ 05:09 0:00
mpirun -am ft-enable-cr --app /home/ccs/
ompi_global_snapshot_13897.ckpt/restart-appfile
Then I checkpointed it:
ccs@grid-demo-1:~$ ompi-checkpoint 14005
When executing this checkpoint command, the running application
directly aborts, even though I did not specify the "--term" option:
----------------------------------------------------------------------
----
mpirun noticed that process rank 1 with PID 14050 on node grid-
demo-1.cit.tu-berlin.de exited on signal 13 (Broken pipe).
----------------------------------------------------------------------
----
ccs@grid-demo-1:~$
Interesting. This looks like a bug with the restart mechanism in Open
MPI. This was working fine, but something must have changed in the
trunk to break it.
A useful piece of debugging information for me would be a stack trace
from the failed process. You should be able to get this from a core
file it left or If you would set the following MCA variable in
$HOME/.openmpi/mca-params.conf:
opal_cr_debug_sigpipe=1
This will cause the Open MPI app to wait in a sleep loop when it
detects a Broken Pipe signal. Then you should be able to attach a
debugger and retrieve a stack trace.
The "ompi-checkpoint 14005" command however does not return.
Is anybody here using checkpoint/restart capabilities of OMPI?
Did anybody encounter similar problems? Or is there something
wrong about my way of using ompi-checkpoint/ompi-restart?
I work with the checkpoint/restart functionality on a daily basis,
but I must admit that I haven't worked on the trunk in a few weeks.
I'll take a look and let you know what I find. I suspect that Open
MPI is not resetting properly after a checkpoint.
Any hint is greatly appreciated! :-)
Best,
Matthias
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users