Design And Implementation Of Scalable Performance Simulator For Exascale Computing

With the performance of parallel computer systems forwarding to exascale frompetascale, the performance evaluation using simulation techniques is becoming partic-ularly important in the design and implementation of a large-scale parallel computer sys-tem. However, the huge scale of the system and great complexity of the application havebrought new challenges to the performance simulation technology of exascale systems.On the one hand, the scalability of the existing simulation techniques is limited, thus it'snecessary to design a new performance simulator which can be extended to the exascalesystems. On the other hand, the performance simulation of large-scale systems has putforward higher requirements of accuracy and efficiency, so new optimization techniquesto improve the simulation accuracy and efficiency need to be developed.Based on the existed research of performance simulation, this study mainly containsthree aspects: the design, optimization and implementation of the scalable performancesimulator.Firstly, we proposed a new design of scalable performance simulator: EDPVC-SIM.Itisanexecution-drivensimulatorbasedonvirtualprocessclock,andthebasicideasofthedesign contain simulating process by process and execution-driven simulation based onprocessvirtualclock. IntherunningprocessoftheEDPVC-SIMsimulator,thesequentialexecutiontimeisobtainedbytheactualexecutionoftheprogram, andthecommunicationoverhead is calculated by simulation. The simulator maintains respective virtual clocksfor all the processes, which are synchronized by message passing.Secondly, for the optimization of the scalable simulator, this paper took account ofboth the simulation accuracy and efficiency to optimize the simulator. Firstly, we pro-posed a basic method to optimize the simulation accuracy after analyzing the precision ofobtaining sequential execution time. Then, on the basis of the communication model ofthe LogP family, we proposed a new model based on the characteristics of the commu-nication protocol, which optimized the accuracy of communication overhead calculation.At last, the paper focused the research on task mapping algorithms which have a signif-icant impact on the simulation efficiency, and a heuristic task mapping algorithm basedon simulated annealing was presented.Lastly, we completed the prototype system of the scalable EDPVC-SIM simulator. Based on the analysis of the structure and implementation principle of MPICH2 library,thispaperproposedtheimplementationframeworkofEDPVC-SIM,andthenimplement-ed the basic point-to-point communication interfaces and common group communicationinterfaces for performance simulation. The simulation accuracy, simulation efficiencyand task mapping algorithm were included in our experimental tests, which finally veri-fied the correctness and effectiveness of the performance simulation technology proposedin this paper.