Toward auto-scheduling compilers

Abstract

In this paper we propose a general framework for compiling, scheduling, and executing parallel programs on parallel computers. We discuss important aspects of program partitioning, scheduling, and execution, and consider possible realistic alternatives for each issue. Subsequently we propose a possible implementation of an auto-scheduling compiler and give simple examples to illustrate the principles. Our approach to the entire problem is to utilize program information available to the compiler while, at the same time, allowing for run-time “corrections” and flexibility.