Design And Implementation Of Discrete-Event Execution-Driven Performance Simulator Based On SMP

The progress of large scale applications for biomolecular simulation,aerospace com-puting, hurricaneforecast,andsoon, promotedthecontinueddevelopmentofhighperfor-mance computing systems in recent years.In order to reduce costs, the system developersneed to conduct a comprehensive performance analysis before the completion, to eval-uate whether the applications can achieve the expected performance, and then to guidethe system design,hardware and software selection, and to provide reference for the im-provement and optimization of the design and implementation of the system. Simulationmethod is a mainstream technology for parallel system performance analysis and predic-tion,due to the accuracy,time cost and flexibility advantages compared with other meth-ods, it has been widely used.In this paper, we focused on the research of simulation technology, the main contentsinclude the design,implementation and experimental verification of the discrete-eventexecution-driven performance simulator.Firstly, we analyzed the key issues for designing of SMP-based discrete-event simu-lator,then put forward the corresponding solutions.The core idea of the SMP-SIM simu-lator includes two aspects, one is the global management of the event with the granularityof CPU,and global data structures are stored in the shared memory.The basic idea is thatevents with smallest time are always been chosen to process;The other is combination ofvirtual and actual execution for event processing: When an event for communication isselected for processing, virtual processing will be used and the communication overheadwill be computed by communication model.If the event is serial computing event,then itwill be actually executed.Secondly,weproposedthedesignschemeoftheSMP-baseddiscreteeventsimulator:SMP-SIM, and analyzed the principle of the key function modules in detail.In SMP-SIM,thesimulationAPIlayerwasintroducedintotheMPIlibrarytoimplementthemainsimu-lationfunction,whichcontainsthreefunctionmodules: communicationprimitivesdecom-position module,communication model module and event management module. In com-munication primitives decomposition module, all of the primitives are re-implementedthrough several core peer-to-peer communication primitives.When the target program ex-ecutes a communication primitive on the host, the primitive will be decomposed into core primitives by the decomposition module, then to call these core peer-to-peer communica-tion primitives.These call and return will reduce events, and event managment module isresponsible for the event scheduling,processing, and updating the corresponding virtualtime.Event management module will interact with communication model when commu-nication events are processing, and it’s also responsible for calculating of ralated eventtime.Finally, we implemented the prototype system of SMP-SIM based on the MPICH2communicaiton library, and conducted comprehensive testing and analysis of accurary,efficiency and scalability.