A code isolator: isolating code fragments from large programs

In this paper, we describe a tool we have developed called a code isolator. We envision such a tool will facilitate many software development activities in complex software systems, but we are using it to isolate code segments from large scientific and engineering codes, for the purposes of performance tuning. The goal of the code isolator is to provide an executable version of a code segment and representative data that mimics the performance of the code in the full program. The resulting isolated code can be used in performance tuning experiments, requiring just a tiny fraction of the execution time of the code when executing within the full program. We describe the analyses and transformations used in a code isolator tool, which we have largely automated in the SUIF compiler. We present a case study of its use with LS-DYNA, a large widely-used engineering application. In this paper, we demonstrate how the tool derives code that permits performance tuning for cache. We present results comparing L1 cache misses and execution time for the original program and the isolated program generated by the tool with some manual intervention. We find that the isolated code can be executed 3600 times faster than the original program, and most of the L1 cache misses are preserved. We identify areas where additional analyses can close the remaining gap in predicting and preserving cache misses in the isolated code.