Multi-Core Software

  Section 5 of 9  

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

PERFORMANCE RESULTS

After our compiler was successfully threaded and debugged, we spent some time in tuning its performance. Of particular importance was the choice of thread scheduling. We conducted many experiments with various parallel-loop scheduling policies. From the parallel-loop scheduling schemes supported by OpenMP*, self-scheduling provided the best performance. In addition, we implemented a scheduling policy that consistently outperformed self scheduling. The policy took advantage of the information that the compiler has about the functions it needs to compile. As part of parsing the input file and creating the intermediate language, the compiler has a substantial amount of information about the structure and the size of each function. We used this information as a static estimate of the time it would take to compile each function. We then grouped together functions in as many chunks as the number of threads or available cores in such a way that the workload of each chunk is almost the same. Through this technique we avoided the load imbalance problem. Figure 6 shows the parallel speedup we achieved in comparison to the theoretical speedup limit. The results are based on our experiments on a 4-socket dual-core system–a total of eight processors. We also spent some time in making sure lock contention was reduced by proper choice of locking. We were pleased with the final parallel performance of the threaded compiler as it approached the theoretical limit of parallel performance as dictated by Amdahl's law. Figure 6 shows the speedup of the threaded compiler compared to the original sequential compiler when compiling the SPEC CPU2000 benchmarks.

Figure 6: Parallel speedups of compiling CPU2000 benchmarks
click image for larger view
 

  Section 5 of 9  

Back to Top