Performance Analysis Of OpenMP Parallel Programs

With the development of large scale parallel computing technology in recent years, especially with the continuously development of multi-core technology, the multithreaded programming technology has become more and more popular. OpenMP is the de facto standard in the field of programming on shared memory architecture. Developing parallel program with OpenMP faces some performance problems, such as load imbalance and low memory access efficiency, etc. Some performance analysis tools have been developed to find and locate such performance problems. But they suffer from no cross-platform capability, due to their dependency on operation system or hardware platform. A new light-weight performance analysis tool with cross-platform capability is designed and implemented in this thesis.An overhead analysis model on OpenMP programming is introduced first. And the performance analysis unit is presented, which is used to analyze load balancing. The OpenMP program performance problems based on specification language of performance property is also precisely defined. The implementation details of performance analysis tool are presented in the last of this thesis. The source-level instrumentation is used in this tool to analyze various performance problems defined in the OpenMP program. And the hardware information during program execution is recorded in account registers, a group of special flag registers in processor. The account registers record the execution context, which can be used to analyze some performance characters such as cache missing rate. The analysis tool accesses the account registers by PAPI. The analysis tool can also report the load balancing degree of performance analysis unit and detect the performance degradation in a semi-automation way. This tool, implemented in standard modular using ANSI C, shows its advantages in cross-platform capability and good scalability.The performance analysis tool is verified by using standard benchmark. The performance log of analysis tool can not only locate the performance problems in source code, but also demonstrate the reason of performance problems.