Efficient framework architecture for improved tuning time and normalized tuning time

摘要

To improve the performance of the program by finding and applying the best set of techniques from the set of available techniques for the GCC Compiler is a non-trivial task. GCC Compiler provides three different levels of optimization techniques. Some techniques are turned ON or OFF each time. Turning on all the techniques for an application may degrade the program performance and increase the compilation time. The selection is dependent on the program domain, the compiler setting and the system architecture. The objective is to find the best set of techniques from the various options provided by the GCC compiler. The framework should handle all the new set of techniques that get added with the new release. The framework should be capable of finding the best set of optimization techniques for any program that is provided as input. Since there are more number of techniques, finding the best set is not feasible manually. The process of selection is automated in order to minimize the execution time, compilation time, tuning time and normalized tuning time. The existing algorithms such as the Combined Elimination, Batch Elimination, Branch and Bound and Optimality Random Search are modified and the results are compared with the newly developed Push and Pop Optimal Search Algorithm. We have implemented the algorithms on both a Pentium IV machine and a Core 2 Duo machine, by measuring the performance of MiBench benchmark programs under a set of 65 GCC compiler options. It is found that the Push and Pop algorithm shows better performance when compared with other algorithms.