Modeling Relations between Inputs and Dynamic Behavior for General Programs

Program dynamic optimization, adaptive to runtime behavior changes, has become increasingly important for both performance and energy savings. However, most runtime optimizations often suffer from the lack of a global picture of a program’s execution, and cannot afford sophisticated program analysis. On the other hand, offline profiling techniques overcome both obstacles but are oblivious to the effects of program inputs. An approach in the between is to offline find the connections between program inputs and runtime behavior, and then apply the knowledge to runtime optimizations. Although it potentially gets the best of both worlds, it yet faces an open question: How to discover and model the relations between inputs and runtime behavior for general programs. This work explores that fundamental problem in three aspects. It proposes an eXtensible Input Characterization Language (XICL) to resolve the complexity of program inputs. With XICL translator, a raw input is automatically converted to an attribute vector, which is then refined by a feature selector to remove redundancies and noises. The final step employs statistical learning techniques to build input-behavior models, mainly through cross-validation, regression trees, and polynomial regression. Experiments on IBM XL compilers show accurate prediction of detailed execution profiles, which helps profile-directed compilation outperform both static and offline profiling-based compilations in a commercial compiler, demonstrating the potential of the technique for continuous program optimizations.