Efficient Algorithm for Symmetric Matrix Data Redistribution

Abstract

In this paper, we present an efficient algorithm for BLOCK-CYCLIC data redistribution to minimize communication costs on symmetric matrix. The main idea of the algorithm is to explore all identical elements in local space. By realigning those elements into corresponding destination data layout, inter-processor data transmission overheads can be reduced. Given a BLOCK-CYCLIC (x, *) to BLOCK-CYCLIC (y, *) symmetrical matrix data re-decomposition over P processors, the theoretical analysis shows 1/P data packing/unpacking and communication costs can be reduced. The experimental results reflect small extra computational overheads of the Symmetric Redistribution Algorithm (SRA) will be incurred. However, the SRA technique can still provides superior overall performance.