MPI reduction operations for sparse floating-point data

Abstract

This paper presents a pipeline algorithm for MPI_Reduce that uses a Run Length Encoding (RLE) scheme to improve the global reduction of sparse floating-point data. The RLE scheme is directly incorporated into the reduction process and causes only low overheads in the worst case. The high throughput of the RLE scheme allows performance improvements when using high performance interconnects, too. Random sample data and sparse vector data from a parallel FEM application is used to demonstrate the performance of the new reduction algorithm for an HPC Cluster with InfiniBand interconnects. © 2008 Springer-Verlag Berlin Heidelberg.