Tight Bounds for Low Dimensional Star Stencils in the Parallel External Memory Model

Abstract

Stencil computations on low dimensional grids are kernels of many scientific applications including finite difference methods used to solve partial differential equations. On typical modern computer architectures, such stencil computations are limited by the performance of the memory subsystem, namely by the bandwidth between main memory and the cache. This work considers the computation of star stencils, like the 5-point and 7-point stencil, in the external memory model and parallel external memory model and analyses the constant of the leading term of the non-compulsory I/Os. While optimizing stencil computations is an active field of research, there has been a significant gap between the lower bounds and the performance of the algorithms so far. In two dimensions, this work provides matching constants for lower and upper bounds closing a multiplicative gap of 4. In three dimensions, the bounds match up to a factor of $\sqrt{2}$ improving the known results by a factor of $2\sqrt{3}\sqrt{B}$, where $B$ is the block (cache line) size of the external memory model. For dimensions $d\ge 4$, the lower bound is improved between a factor of $4$ and $6$. For arbitrary dimension~$d$, the first analysis of the constant of the leading term of the non-compulsory I/Os is presented. For $d\ge 3$ the lower and upper bound match up to a factor of $\sqrt[d-1]{d!}\approx \frac{d}{e}$.

64 pages, 8 figures, 4 tables

References in corpus (5)

• Minimizing Communication in Linear Algebra
• Efficient multicore-aware parallelization strategies for iterative stencil computations
• Graph Expansion and Communication Costs of Fast Matrix Multiplication
• Communication-Optimal Parallel Algorithm for Strassen's Matrix Multiplication
• Tight Bounds for Low Dimensional Star Stencils in the Parallel External Memory Model

Cited by in corpus (1)

• Tight Bounds for Low Dimensional Star Stencils in the Parallel External Memory Model