Assist Threads for Data Prefetching in IBM XL Compilers

Processor chips in development today typically support multiple hardware threads of execution. When an application does not exhibit enough parallelism to effectively use all available threads, the extra threads can be used as assist threads to prefetch data for the main thread, and thus improve performance. In our model, the main thread performs all useful work in the application. Work done in the assist thread is not necessary for correct execution of the application, nor does it interfere with any results generated by the application. Thus, we can throttle assist thread execution or skip work in the assist thread in order to synchronize with the main thread. In this paper, we describe the IBM XL compiler transformation that automatically generates prefetching code for the assist thread, and optimizes the resulting multi-threaded code by inserting synchronization. We also describe the runtime system that is used to control execution of the assist thread with respect to the main thread. We present experimental results that show the potential benefit of using assist threads to prefetch data in a system with Power5 processors.

Greg Steffan

Last modified: Wed Aug 26 17:52:21 EDT 2009