[mpich-discuss] Error in MPI_Finalize on a simple ring test over TCP

[Wesley Bland]

The value of previous for rank 0 in your code is -1. MPICH is complaining because all of the requests to receive a message from -1 are still pending when you try to finalize. You need to make sure that you are receiving from valid ranks.

On Jul 10, 2013, at 7:50 AM, Thomas Ropars <thomas.ropars at epfl.ch> wrote:

> Yes sure. Here it is.
> 
> Thomas
> 
> On 07/10/2013 02:23 PM, Wesley Bland wrote:
>> Can you send us the smallest chunk of code that still exhibits this error?
>> 
>> Wesley
>> 
>> On Jul 10, 2013, at 6:54 AM, Thomas Ropars <thomas.ropars at epfl.ch> wrote:
>> 
>>> Hi all,
>>> 
>>> I get the following error when I try to run a simple application implementing a ring (each process sends to rank+1 and receives from rank-1). More precisely, the error occurs during the call to MPI_Finalize():
>>> 
>>> Assertion failed in file src/mpid/ch3/channels/nemesis/netmod/tcp/socksm.c at line 363: sc->pg_is_set
>>> internal ABORT - process 0
>>> 
>>> Does anybody else also noticed the same error?
>>> 
>>> Here are all the details about my test:
>>> - The error is generated with mpich-3.0.2 (but I noticed the exact same error with mpich-3.0.4)
>>> - I am using IPoIB for communication between nodes (The same thing happens over Ethernet)
>>> - The problem comes from TCP links. When all processes are on the same node, there is no error. As soon as one process is on a remote node, the failure occurs.
>>> - Note also that the failure does not occur if I run a more complex code (eg, a NAS benchmark).
>>> 
>>> Thomas Ropars
>>> _______________________________________________
>>> discuss mailing list     discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
>> 
>> 
> 
> <ring_clean.c>_______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss