[mpich-discuss] Errors: In direct memory block for handle type REQUEST, 2 handles are still allocated

[Dave Goodell]

On Oct 26, 2011, at 4:12 PM CDT, Borrelli, Steven wrote:

> I'm seeing the following issues when I attempt to run a program that used to run on 1.2.1p1/mpd.  What should be my next step in debugging this issue? 
> 
> 
> Results for launching mpiexec with -n
> 
> N=1, run completes without any issues
> N=2,3, run completes, but has the "In direct memory block for handle type REQUEST, 2 handles are still allocated" warning

This message indicates that you are leaking requests somehow.  Perhaps you aren't invoking MPI_Wait (or equivalent) on every request in your program?

> N=4+, run fails with a signal 11. 

Hard to say from the included info exactly what's causing the segfault.  I'd try running your application under valgrind.  Something like:

mpiexec -n $NUM_PROCS valgrind -q ./myapp

I'd start with NUM_PROCS=1 and work up from there after fixing any errors encountered at each level.

-Dave