A Performance Tool for High-Level Parallel Programming Languages

Abstract

Users of high-level parallel programming languages require accurate performance information that is relevant to their source code. Furthermore, when their programs experience performance problems at the lowest levels of their hardware and software systems, programmers need to be able to peel back layers of abstraction to examine low-level problems while maintaining references to the high-level source code that ultimately caused the problem. In this paper, we present NV, a model for the explanation of performance information for programs built on multiple levels of abstraction. In NV, a level of abstraction includes a collection of nouns (code and data objects), verbs (activities), and performance information measured for the nouns and verbs. Performance information is mapped from level to level to maintain the relationships between low-level activities and high-level code, even when such relationships are implicit.

We have used the NV model to build ParaMap, a performance tool for the CM Fortran language that has, in practice, guided us to substantial improvements in real CM Fortran applications. We describe the design and implementation of our tool and show how its simple tabular and graphical performance displays helped us to find performance problems in two applications. In each case, we found that performance information at all levels was most useful when related to parallel CM Fortran arrays, and that we could subsequently reduce each application’s execution time by more than half.

This research was supported in part by Department of Energy grant DE-FG02-93ER25176, Office of Naval Research grant N00014-89-J-1222, and National Science Foundation grants CCR-9100968 and CDA-9024618.