Interprocedural communication optimizations for distributed memory compilation

Abstract

Managing communication is a difficult problem in distributed memory compilation. When the exact data to be communicated cannot be determined at compile time, communication optimizations can be performed by runtime routines which generate schedule for communication. This leads to two optimization problems: placing communication so that data once communicated can be reused if possible and placing schedule calls so that the result of runtime preprocessing can be reused for communicating as many times as possible. In large application codes, computation and communication is spread across multiple subroutines, so acceptable performance cannot be achieved without performing these optimizations across subroutine boundaries. In this paper, we present an Interprocedural Analysis Framework for these two optimization problems. Our optimizations are based on a program abstraction we call Control & Call Flow Graph. This extends the call graph abstraction by storing the control flow relations between various call sites within a subroutine. We show how communication placement and schedule call placement problems can be solved by data-flow analysis on Control & Call Flow Graph structure.

This work was supported by NSF under grant No. ASC 9213821 and by ONR under contract No. N00014-93-1-0158. The authors assume all responsibility for the contents of the paper.