Visual Parallel Performance Debugging Environment: Design And Implementation

Currently most parallel machines achieve poor performance, it is becausea lack of effective parallel program development tools, and the poor functionof current debug tools can't meet the need. These make the programming,debugging and optimizing of parallel program difficult. The focus of thisthesis is Message-Passing parallel program, which is more flexible, moreuniversal than other parallel program model, and is widely used.The major work of this thesis is as below:1) Improve the parallel performance debugging model. Instrument — collectperformance data — modify the instrumentation code — collect performance data.Visualize and analyze the data, modify the source code, until the bottleneck is found,and corrected.2) Improve the physical clock synchronization model which is adapted forparallel performance debugging environment. New model highly improves theaccuracy of adjustment of local clock. Its error is less than 15% of previous one's. Itimproves the stability and precision when the scale increases. And, the new modelkeeps no introduce any communication overhead. Its time cost is reasonable.3) Modified MPI source code, the MPI program will run under the control ofdebugger, which makes it convenience for debugger to control the MPI program andcollect performance data.4) Implement dynamic code instrumentation by using Dyninst tools to collectperformance data and debugging information. The instrumentation code can be usedto collect source code information, debugging information and performance data, andgather information of the sources' utilization.5) Analysis the data, assemble the main performance data and debugginginformation into three tables. Clock information table records the information offunction call, function information table records the source code information of everyfunction, and the call statistics of the functions, communication information tablerecords communication events.6) Based on the process time-space chart, design the proceduretime-space chart. The events' information and procedure call information canbe check on the chart. The time-space chart will have the source codeclick-back and click-forward function, and will present the critical path ofparallel program.