Harmony-based Performance Evaluation Method For Dynamic Multi-version

With the increases of hardware complexity and the participation of virtualization in layers of execution stack, prediction of performance impact of various optimizations becomes more and more difficult. Modern compilers and virtual machines spend substantial money and efforts in performance tuning in order to achieve good performance for a range of benchmarks. Optimizations normally present good average performance improvement, but they often introduce significant performance degradation because of unpredictable impact at runtime. Modern VMs utilize complicate feedback-directed optimizations to address performance degradations, but they usually make the situation worse by making the compiling system more unpredictable.This paper describes a new algorithm to evaluate the effectiveness of optimizations and audit, correct the performance degradations based on an open source JVM, DRLVM in Harmony Framework. The empirical results show that the work achieved a good performance improvement compared to traditional compiling strategy. The average improvement of performance for all benchmarks used in this project increases from 2.250% to 2.648%, and the ratio of benchmarks with performance degradation because of continuous compilation drops to nearly 20.0% compared to the original 33%.