SOFTWARE: Heterogeneous Cluster MPI: A System for Out-Of-Core, Heterogeneous Data Distribution

University Of Georgia Research Foundation Inc, Athens GA

Investigators

Abstract

Grid computing, or using several potentially heterogeneous computers to execute an application, is becoming increasingly popular. A grid poses significant software challenges for parallel computing; in particular, fundamental parallel computing problems are made more difficult. This project focuses on the data distribution problem, which is to distribute data to processors to minimize application completion time. We will re-examine the data distribution problem on one specific instance of a grid: a local area network of heterogeneous computers. Processors in such a cluster can have different relative CPU speeds, different amounts of available physical memory, and different effective I/O latencies. In addition, we will support the increasing number of out-of-core programs on these clusters. We will address this data distribution problem through the design and implementation of what we call Heterogeneous Cluster MPI (HC-MPI), which will be constructed as extensions and analysis built into MPI. The extensions will (1) assist users in writing adaptable programs and (2) allow HC-MPI to collect enough program information to determine automatically an effective data distribution. The expected impact of this work is to allow computational scientists to execute parallel programs efficiently on a variety of heterogeneous clusters.