Multi-Core Software

Methodology, Tools, and Techniques to Parallelize Large-Scale Applications: A Case Study

INTRODUCTION

In this paper we focus primarily on the techniques used to parallelize an application, the tools that facilitate the parallelization, and the new insights this approach yielded. Techniques that proved helpful in our work are at the core of a comprehensive solution suite Intel is developing to assist software developers discover and exploit parallelism in their applications. The generally applicable source changes necessary to make the compiler thread safe are also categorized and described. As expected, good software engineering principles such as modularization, data abstraction, and information hiding ease the process of threading an application. We also describe how we automated repetitive source changes. To make it feasible to apply all the source changes necessary for an application of this size, where the threaded loop spans hundreds of modules covering hundreds of thousands of lines of code with extensive use of macros, semi-automated script tools were developed. It is easy to get overwhelmed by the data dependence complexity and size when starting to thread existing serial applications, but as we hope to illustrate in this case study, with the help of Intel's threading tools and a systematic approach, it is possible to achieve large application threading with a reasonable amount of effort and time.

The threading effort, involving a small team over a relatively short period of time, successfully yielded a working parallelized compiler. Although work remains to be done in tuning the resulting application, we also discuss in this paper the impact of different thread scheduling algorithms and the speedups achieved. We also briefly discuss the issues involved in maintaining a thread-safe application.

Back to Top