| Author | W. K. Kaplow, B. K. Szymanski, P. Tannenbaum and V. K. Decyk |
|---|---|
| Title | All Cache Virtual Machine: A Method for Improving the Cache Performance of Irregular Scientific Codes |
| Year | 1995 |
| Abstract | In this paper we present the All Cache Virtual Machine (ACVM) method for improving the cache performance of irregular scientific codes. The method divides a real processor into a set of virtual machines based on the real processor cache size, and computation size and data reference pattern. Data references made to arrays distributed over ACVM's are rearranged to improve cache reuse. For dynamic irregular codes there is a run-time component that redistributes these data to maintain the cache reuse high. This run-time component is copied from the load balance code which is part of the original program and which maintains data locality in the main memory for dynamic irregular parallel programs. A novel set of compiler directives is introduced that can be used by the programmer to indicate to a compilation system the data structures to which the ACVM method should be applied. A preliminary directive preprocessor is described. Results of the ACVM method are shown on Particle-In-Cell (PIC) codes that are used to simulate plasma physics, and for sparse matrix-vector multiplication which is at the heart of many iterative solvers of linear equations. Performance results for execution on several processor types are also presented. |
| PDF File |  Download |