[mvapich-discuss] Seg fault in MPI_Finalize with many windows

[Pavan Balaji]

On 10/08/2010 05:55 PM, sreeram potluri wrote:
> This is related to how ib registration information for the windows is
> managed. I see that this limitation can be addressed without making any
> sub-optimal choices. We plan to fix this in the upcoming release.

Ok, I looked through the code. You are storing the registration 
information in a global structure outside the window 
(MPIDI_CH3I_RDMA_Process). This is against the MPICH2 coding style. All 
information related to the window should reside in the MPID_Win 
structure (or more specifically, an internal device-specific handle such 
as MPID_DEV_WIN_DECL).

Would you be able to give us a patch soon-ish? As Jim mentioned, it's 
hard to figure out how many windows the application would create. (I 
guess we could give a very large number, but that sounds like an awful 
hack).

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji