SparseBLAS Products in UPC: an Evaluation of Storage Formats

The performance of a significant number of applications in High Performance Computing (HPC) is determined by the efficiency of the sparse matrix-vector and matrix-matrix products. These computational kernels generally present poor scalability due to the lack of memory local-ity exploitation. Selecting the most appropriate storage format, which generally depends on the specific application scenario, can significantly improve their efficiency. This paper presents an evaluation of the most common sparse storage formats using Unified Parallel C (UPC). UPC is a Partitioned Global Address Space (PGAS) language that provides high programmability and per-formance through an efficient exploitation of data locality, especially on hierarchical architectures such as multicore clusters. Different combinations of storage formats and data distributions for the SparseBLAS matrix products are analyzed. Experimental results on an HP supercomputer using representative sparse matrices show that a suitable combination of storage formats and parallel algorithms has a great influence on the performance of the sparse products