[mpich-discuss] Crash with hydra on MPI_Finalize

[Dave Goodell]

Crash how?  Is it segfaulting?  Is there an error message that you can show us?

The MPIR_CommL_forget function only gets called when debugger support is enabled, so there's a chance you're hitting a bug there that others do not because many (most?) users do not enable debugger support.  If you don't need debugger support (for inspecting message queues and the like, not plain-old "-g" compilation), then you could remove "--enable-debuginfo" from your configure line as a workaround.

Based on the information you've given us so far, I doubt this is a crash that is caused by hydra vs. mpd.  Did you enable debugger support when building your older version of MPICH2 as well?

-Dave

On Mar 5, 2011, at 2:09 AM CST, Jain, Rohit wrote:

> Hi,
>  
> I am seeing a crash with mpich2-1.3.3rc1, when application quits. This is a specific case where all processes are exiting, and they call MPI_Finalize almost at the same time.
>  
> # 0    0x08c2bff8: 'MPIR_CommL_forget + 0x28'
> # 1    0x08c1ff52: 'comm_delete + 0x1c2'
> # 2    0x08c20145: 'MPIR_Comm_release_always + 0x35'
> # 3    0x08c5358e: 'MPID_Finalize + 0x2e'
> # 4    0x08c27089: 'MPI_Finalize + 0x119'
>  
> Note that I recently started using hydra with this mpich2 version. Earlier mpich2 version with mpd works fine.
>  
> Is there any difference in how hydra handles finalize?
> If so, does application need to change to make it work with hydra?
>  
> Regards,
> Rohit
>  
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss