Generating parallel algorithms for cluster and grid computing

Abstract

We revisit and use the dependence transformation method to generate parallel algorithms suitable for cluster and grid computing. We illustrate this method in two applications: to obtain a systolic matrix product algorithm, and to compute the alignment score of two strings. The product of two n × n matrices is viewed as multiplying two p × p matrices whose elements are n/p × n/p submatrices. For m such multiplications, using p2 processors, the proposed parallel solution gives a linear speedup of mp 3/(m+2)p-2 or roughly p2. The alignment problem of two strings of lengths m and n is solved in O(p) communication rounds and O(mn/p) local computing time. We show promising experimental results obtained on a 16-node Beowulf cluster and on an 18-node grid called InteGrade, consisting of desktop computers. © Springer-Verlag Berlin Heidelberg 2005.