A Parallel Accelerated Method For Dynamic Analysis And Profiling Based On Multi-cored Environment

Dynamic instrumentation system injects code into the application when it is executed to gain dynamic profiling data or extend the application's behavior. It is widely used in program profiling and analysis, architectural simulation and bug detection, etc. However, dynamic instrumentation systems suffer from performance problems. When do dynamic analysis especially online profiling of long-time running programs. On the other hand, multi-core and many-core processors are becoming more and more prevalent in modern computing environments. They provide more parallelism for software to exploit.We present the design and implementation of MT-Profiler, which is a parallel dynamic analysis framework based on our two-stage sampling strategy. It introduces a two-stage sampling strategy to perform low overhead dynamic analysis. It also provide several optimizations to instrument efficient checking code and create slice which is instrumented and executed analysis code parallel with application thread to exploit threaded parallelism. In addition, MT-Profiler presents a strategy to monitor system's real-time load, uses the monitoring information to control system load balance and feedback two-stage sampling strategy to adjust the sampling rate. Finally, MT-Profiler provides a set of easy to use, efficient API for tool writers to use so that parallel speedup their tools.We evaluated the MT-Profiler system across NPB3.3 OPENMP test suite. The result demonstrates that MT-Profiler is 3 to over 17 times speedup and the accuracy is over 80% on average compared with the DynamoRIO perfect profile. Our experiments also show that MT-Profiler is very tunable in both sampling stages and adjust the sample threshold can gain higher accuracy.