Advanced MPI Support of Distributed Execution of R Programs

• Activity Number: 354
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Chen Ding*+ and Bin Bao and Xiaoming Gu
• Companies: University of Rochester and University of Rochester and University of Rochester
• Keywords: distributed computing ; R ; MPI

Abstract

A modern computer cluster is capable of performance tens or hundreds of times higher than that of a personal computer. To use a cluster, a program must explicitly communicate share data between machines. MPI is widely used in scientific computing, with portable and optimized support for communication over commodity networks. An R package, Rmpi, has recently been developed as a wrapper to MPI. In this paper, we present several improvements to MPI in support of statistical computing in R. For example, it permits an application to compute on a large vector and communicate it at the same time, forming a pipeline. Such pipelines may be dynamically chained at run time, and a program can gain performance more than what is possible with basic computation-communication overlapping. Our results show that this can improve the parallel performance on 30 cluster nodes by integer factors.