[mpich-discuss] mpich collectives topology aware?

Dave Goodell goodell at mcs.anl.gov
Mon Nov 29 08:20:02 CST 2010


Many of our collectives are topology aware, but for some operations there are no straightforward uses of topology.  And some others aren't implemented due to lack of demand and lack of developer bandwidth.  I think that these operations take intranode/internode communication into account:

MPI_Bcast
MPI_Reduce
MPI_Allreduce
MPI_Scan
MPI_Barrier

-Dave

On Nov 24, 2010, at 2:32 PM CST, Darius Buntinas wrote:

> 
> Yes, I believe all collectives are topology aware.  I checked Allreduce specifically, and it does what you describe.
> 
> -d
> 
> 
> On Nov 24, 2010, at 1:29 PM, Filip Blagojevic wrote:
> 
>> Hi,
>> 
>> Could anyone tell me if MPI_Allreduce() (and other collectives) is topology aware?
>> More specifically, I would like to know if MPI_Allreduce() will first perform a local reduction (within a single shared memory node), and then only one MPI rank will be chosen on each node to perform inter-node reduction. Or maybe the reduction is not aware of what MPI processes are sharing a node?
>> 
>> Thank you,
>> Filip_______________________________________________
>> mpich-discuss mailing list
>> mpich-discuss at mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
> 
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss



More information about the mpich-discuss mailing list