Improved Internode Communication for Tile QR Decomposition for Multicore Cluster Systems

Tile algorithms for matrix decomposition can generate many fine-grained tasks. Therefore, their suitability for processing with multicourse architecture has attracted much attention from the high-performance computing (HPC) community. Our implementation of tile QR decomposition for a cluster system has dynamic scheduling, OpenMP work- sharing, and other useful features. In this article, we discuss the problems in internodes communications that were present in our previous implementation. The improved implementation has both strong and weak scalability.