On the Efficient Implementation of Reductions on the Cell Broadband Engine

For a high-performance parallel implementation of many scientific algorithms, efficient realizations of combining communication patterns like reduce or all-reduce are important. Especially on the Cell Broadband Engine a low latency realization of such operations is not obvious. So in this paper several algorithms for implementing reductions are discussed and efficient implementations on the Cell are proposed. Detailed performance results are presented for reductions of vectors of various sizes on a Cell blade consisting of two interconnected Cell processors. It is shown that the new reductions algorithms are in most cases faster than other previously published implementations.