MPICH-2-MX - Attach to Proces

[CorCornelisse]

I'm in the process of evaluation the Totalview debugger and am experiencing a little difficulty when trying to attach to the debug process.

I'm using MPICH2-MX (which is a MPICH2 version patched by Myrinet), I'll post the output of mpich2version:
============
MPICH2 Version: 1.0.6
MPICH2 Release date: Unknown, built on Tue Feb 26 15:13:57 CET 2008
MPICH2 Device: ch_mx
MPICH2 configure: --prefix=/home/ccorneli/local/mpich2 --with-device=ch_mx --with-mx=/usr/local/Cluster-Apps/mx/mx-1.2.0j-2.6.18-clustervision-136.1_cvos-no-compat-syms --with-pm=smpd --with-pmi=smpd --with-mpe --with-romio --enable-compiler-optimizations --enable-cxx --enable-debuginfo --enable-fast --enable-g=dbg --enable-sharedlibs=gcc --enable-totalview
MPICH2 CC: pathcc -march=opteron -m64 -fno-second-underscore -g -O2
MPICH2 CXX: pathCC -march=opteron -m64 -fno-second-underscore -O2
MPICH2 F77: pathf95 -march=opteron -m64 -fno-second-underscore -O2
MPICH2 F90: pathf95 -march=opteron -m64 -fno-second-underscore -O2
MPICH2 Patch level: none

============

I use the SMPD process manager instead of the MPD manager due to its SGE integration. The problem is, once I attach to one of my processes, it works, but somehow it fails to find the slave processes. My job runs on 29 nodes, each node runs 2 processes, so I have a total of 58 processes. As mentioned before I use the SMPD process manager, I tried attaching to the parent SMPD process as well as the individual processes on the start node (the node on which I issue the mpiexec command). The mpiexec version I have does not appear to have the tvsu or tv command line options.

Is Totalview process attaching not supported this way? (using SMPD) Should I switch to MPD or can it be something else?

[Josh-TotalView-Tech]

Did you see this post: http://forum.totalviewtech.com/cgi-bin/gforum.cgi?post=376#376. Does it help?

[CorCornelisse]

Hi,

I've read the topic, but the problem is:

-My mpiexec isn't a python script (Maybe because of SMPD ? Instead of MPD?)
-I have no tvsu or tv options. It tells me "unrecognized parameter" and just ignores the flag

So, should I reinstall and configure MPICH2-MX to use MPD instead of SMPD? Then I would assume TotalView doesn't support the SMPD process manager?

[Josh-TotalView-Tech]

Can you please report this to support@totalviewtech.com. Please refer to this post when you do so. When we figure this out we will post the result.

Thanks

[CorCornelisse]

Well here I'm again, bit late, since I forgot to post the results. I did some more testing, and TotalView Debugger can only attach if you use MPDs (the python version of MPICH), so, process attachment to running processes on a smpd basis is not working. This was tested using MPICH2 and TotalView Debugger version 8.4 (evaluation license).

Hope this might help somebody (would be nice to include this in the Total View installation instructions)[cool]