Thanks I will pass that on to the person asking. And it answers my
own questions of observed behavior with IMB and other tools.
Center for Advanced Computing
On Dec 10, 2008, at 10:37 AM, George Bosilca wrote:
> Open MPI has a very fancy collective implementation. While we don't
> use multicast, we dynamically change the algorithm used for the
> collective based on the number of nodes, the size and type of the
> message, and eventually the operation involved if any. Moreover, we
> detect hierarchies, and we can potentially change the algorithms to
> take advantage of this knowledge.
> If you want more information, please refer to this paper (Flexible
> collective communication tuning architecture applied to Open MPI)
> for more information, http://www.netlib.org/netlib/utk/people/
> On Dec 9, 2008, at 21:49 , Brock Palen wrote:
>> Is OpenMPI using Multicast for things like MPI_Bcast() ?
>> Some simple tests of really old versions (1.0.2) showed binary
>> tree type performance, while latter versions gave me the same run
>> time no matter how many CPU's I used.
>> Thus this lead me to believe that something like multicast is in
>> use. Is it?
>> Brock Palen
>> Center for Advanced Computing
>> users mailing list
> users mailing list