The paper proposes a novel approach for optimizing performance of All-to-All collective communication by taking advantage of concurrency available in modern networks such as Infiniband or Quadrics. Using as an example MPI_AllGather operation we describe how network concurrency can be exploited in optimized implementation of this operation. For example, comparing to the performance of this operation in leading MPI implementations, for 32KB message size on 128 processors, the new algorithm yields 65% improvement on the Infiniband at Virginia Tech and 89% on the Quadrics cluster at PNNL.
Revised: September 5, 2007 |
Published: June 1, 2005
Citation
Tipparaju V., and J. Nieplocha. 2005.Optimizing All-to-All Collective Communication by Exploiting Concurrency in Modern Networks. In Supercomputing 2005, ACM/IEEE Conference, 46. Los Alamitos, California:IEEE Computer Society.PNNL-SA-45780.doi:10.1109/SC.2005.46