[mpich-discuss] MPICH2 (or MPI_Init) limitation | scalability

Darius Buntinas buntinas at mcs.anl.gov
Wed Jan 18 12:04:32 CST 2012


Hmm.  Is eth2 available on all machines?

-d


On Jan 18, 2012, at 8:39 AM, Bernard Chambon wrote:

> Hello,
> 
> Le 12 janv. 2012 à 18:07, Darius Buntinas a écrit :
> 
>> Great, I'm glad it worked.  BTW, the kernel.shmall et.al. parameters are used for System V shared memory, so they'll have no effect when using mmap-ed shared memory, which is how MPICH is configured by default.
>> 
>> -d
>> 
> 
> Now that your patch has solved the shared memory failure, I encountered a new failure when specifying -iface option
> 
> With a minimal piece of code (*) , 
>  with -iface, I always get assertion error like this one :
> >mpiexec -iface eth0 -n 150 bin/basic_test
> [mpiexec at ccwpge0061] control_cb (./pm/pmiserv/pmiserv_cb.c:215): assert (!closed) failed
> [mpiexec at ccwpge0061] HYDT_dmxu_poll_wait_for_event (./tools/demux/demux_poll.c:77): callback returned error status
> [mpiexec at ccwpge0061] HYD_pmci_wait_for_completion (./pm/pmiserv/pmiserv_pmci.c:181): error waiting for event
> [mpiexec at ccwpge0061] main (./ui/mpich/mpiexec.c:405): process manager error waiting for completion
> 
>  without -iface, it works always
> > mpiexec -n 150 bin/basic_test   : it's OK 100 % of time ( without -iface)
> 
> I'm confused with that because I can't provide the 10Gb/s (-iface eth2) to our customers 
> 
> Best regards
> 
> 
> (*) minimal code :
>   if (MPI_Init(NULL, NULL) != MPI_SUCCESS ) {
>    printf("Error calling MPI_Init !!, => exiting \n") ; fflush(stdout);
>    return(1);
>   } else {
>    MPI_Finalize();  
>    printf("It's OK \n") ; fflush(stdout);
>    return(0);
>   }
> 
> 
>> On Jan 12, 2012, at 5:31 AM, Bernard Chambon wrote:
>> 
>>> Hello,
>>> 
>>> Good news, It works !
>>> Whith your patch I can run 255 tasks and perhaps more, without any special
>>> configuration on the machine (*)
>>> 
>>> 
>> 
> 
> ---------------
> Bernard CHAMBON
> IN2P3 / CNRS
> 04 72 69 42 18
> 
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss



More information about the mpich-discuss mailing list