On Fri, 16 May 2003, kok junlee wrote:
> I wish to study the the comunication complexity for the AllReduce function
> provided by mpi... is the performance affected by the lenght of arrays and
> the numbers of processors involve? is there any mathematical function model
> the relation for those variables...? ... :-(
First, a quick note - the MPI standard specifies the semantics of
MPI_ALLREDUCE, but not how it is implemented. The computational
complexity of the operation is probably (although I really haven't thought
about the problem at all) implementation dependent. Certainly, the
performance characteristics of the call is implementation dependent.
The MPI_ALLREDUCE provided by LAM/MPI is actually not too bad to follow -
you might want to take a look at the implementation if you need to find
computational complexity of our algorithm. The function is defined in
<top_src_dir>/share/mpi/allreduce.c for any of the LAM 6.5.x releasesand
uses basic send/recv functions for implementation.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
