On a distributed shared memory (DSM) system, the optimization of memory access is very important for achieving good performance. The authors propose an optimizing compiler which controls a software cache system implemented on a DSM. The software cache consists of a static part related to the compiler and a dynamic part related to cache-managing runtime routines. The compiler controls the static part of the software cache by using information from static analysis. For applications whose behavior can only be dynamically determined, the compiler uses the dynamic part of the software cache. They also propose the application of RISC-oriented optimization techniques to parallel applications on their software cache system. They evaluate the efficiency of the compiler and RISC-oriented optimization techniques on the CM-5 distributed parallel machine. The results show that the compiler and the optimizations considerably improve the performance of basic linear algebra routines: matrix multiplication, Cholesky decomposition and Gaussian elimination.
展开▼
