MT-Profiler: a parallel dynamic analysis framework based on two-stage sampling

  • Authors:

  • Zhibin Yu;Weifu Zhang;Xuping Tu

  • Affiliations:

  • School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China;School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China;School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China

  • Venue:
  • APPT'11 Proceedings of the 9th international conference on Advanced parallel processing technologies
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Dynamic instrumentation systems offer a valuable solution for program profiling and analysis, architectural simulation, and bug detection. However, the performance of target programs suffers great losses when they are instrumented by these systems. This issue is mainly caused by the resource contention between the target programs and the instrumentation systems. As multi-core processors are becoming more and more prevalent in modern computing environments, more hardware resource parallelism is provided for software to exploit. In this paper, we propose to leverage the abundant computing resources of multi-core systems to accelerate program instrumentation. We design and implement a Multi-Threaded Profiling framework named MT-Profiler to dynamically characterize parallel programs. The framework creates instrumenting and analysis code slices which can run along with application threads in parallel on different cores. To further reduce the overhead of dynamic instrumentation, MT-profiler employs two-stage sampling scheme with several optimizations. It is implemented on DynamoRIO which is an open-source dynamic instrumentation and runtime code manipulation system. The performance of our MT-Profiler is evaluated by using the NPB3.3 OPENMP test suite. The results demonstrate that the MTProfiler obtains 3 to 17 times speedup compared with the DynamoRIO perfect profiler while the average accuracy is over 80%.