Abstract:

This paper presents the design and implementation of an OpenMP runtime optimization system based on valueprofiling, called CCRG OpenMP, where the para  llel regions of OpenMP programs can be well optimized based on the common combinations of values. Moreover, only the code within the parallel regions ne eds to be re-compiled and managed at runtime. A distinguished feature of CCRG OpenMP is that a dynamic re-compile technique is adopted to overcome the s  hortcomings of the static muhi-versional approach. A separate dynamic optimizer is developed to collect and analyze value-profiles for reducing the overhead introduced by re-compile. The SPEC OpenMP benchmarks show the presented value-profile based optimization for OpenMP can improve the performance sig nificantly.

Key words: OpenMP, value-profile, runtime optimization