Data-Driven execution of the Tile LU Decomposition

Abstract

The objective of this paper is to analyze, develop and evaluate the tile LU Decomposition using the FREDDO framework. FREDDO is a C++ framework, based on the DDM model of execution, that supports efficient data-driven execution on conventional processors. The performance evaluation shows that FREDDO scales well and tolerates scheduling overheads and memory latencies effectively. The LU implementation is evaluated in both single-node and distributed execution environments. In both cases our framework achieves very good speedups, especially in the larger problem sizes. Particularly, our framework achieves up to 97% of the maximum possible speedup on a single-node and up to 90% of the maximum possible speedup on a 4-node cluster with a total of 128 cores.