A Replay System for Performance Analysis of Multi-Threaded Programs

In recent years, it is a hotspot for program analysis to detect performance bugs in multi-threaded applications. However, traditional record/replay systems focusing on concurrent anomalies have many limitations to tackle the issues of performance bugs, such as replay overhead and imprecision of replay-based execution time. To cope with the problems above, this paper proposes an improved replay system PerfPlay which can be used for the performance analysis of multi-threaded programs. To be specific, we first collect and analyze the requisite information for the program performance. Secondly, the different replay strategies are discussed and then we present a novel schedule-driven strategy to ensure the performance fidelity of replay system. Finally, we study the classical performance problem of “inter-thread unnecessary lock contention” under the framework of PerfPlay. Compared with the traditional replay strategies, our experimental results demonstrate the performance fidelity of PerfPlay. Through the case study, we find a few performance bugs in real-world and further verify the effectiveness of PerfPlay.