Year: 2015
Type: preprint
Abstract: In this paper, we propose an optimization selection methodology for the ubiquitous sparse matrix-vector multiplication (SpMV) kernel. We propose two models that attempt to identify the major performance bottleneck of the kernel for every instance of the problem and then select an appropriate optimization to tackle it. Our first model requires online profiling of the input matrix in order to detect its most prevailing performance issue, while our second model only uses comprehensive structural features of the sparse matrix. Our method delivers high performance stability for SpMV across different platforms and sparse matrices, due to its application and architecture awareness. Our experimental results demonstrate that a) our approach is able to distinguish and appropriately optimize special matrices in multicore platforms that fall out of the standard class of memory bandwidth bound matrices, and b) lead to a significant performance gain of 29% in a manycore platform compared to an architecture-centric optimization, as a result of the successful selection of the appropriate optimization for the great majority of the matrices. With a runtime overhead equivalent to a couple dozen SpMV operations, our approach is practical for use in iterative numerical solvers of real-life applications.
Source: arXiv (Cornell University)
Institution: National Technical University of Athens
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2228597387
