Evaluating Communication Performance Measurement Methods for Distributed Systems

T.T. Le (USA)


Grid Computing, Parallel Applications, Performance Evaluations, MPI Communication.


Grid computing technology is increasingly drawing the attention of computational scientists and engineers around the world. With the current implementation of Globus and MPICH-G2, MPI-based parallel applications are theoretically able to use the grid as one type of parallel computational resources. Practically, making MPI-based parallel applications running effectively on a grid environment requires lots more programming efforts due to low parallel performance of the grid. This is because of the heterogeneity of the grid computational resources and network structure, the low communication quality of the grid network, and the randomness in network traffic loads as well as the structure of communication links between nodes. This paper describes our experiments and analysis that demonstrate issues in measuring MPI communication performance for both parallel systems and computational grids. We applied traditional communication measurement methods to a parallel system with multi level network architecture. The measured results were then examined and analyzed. Based on the results, we demonstrate our arguments that the current measurement methods are not able to apply to computational grids. There is a need to develop new methods that can measure individual node-to-node collective communications with pipelining effects and with minimum measurement overheads.

Important Links:

Go Back