Research On Parallel Simulation Supporting Technologies Based On Multicore CPU And Many Integrated Core Platform

It has become an important way to implement complex system simulations with the help of high performance computing platforms and parallel simulation technology.At present,computing platforms based on multi-core CPU(CMP)and Many Integrated Core(MIC)processors show powerful computing potential,which brings new opportunities to accelerate simulation execution.However,existing parallel simulation supporting technologies are mainly based on multi-processor machines,cluster,CMP+GPU platforms,which are not able to take advantages as abundant threads,local shared memory and collabarative computing between the CMP and MIC platforms.Therefore,doing research on parallel simulation supporting technologies based on CMP+MIC platforms has important theoretical significance and practical value on taking advantages of the new high performance computing resources,meeting the growing demand for calculation simulation,and promoting the spread use of high performance computing platform in the field of simulation.This thesis mainly focuses on the problems when apply current parallel simulation supporting technologies on the CMP+MIC platforms.Deep researches have been carried out on issues such as an architecture of parallel simulation engine,the computing models acceleration and load balancing.Main work and innovations are as follows:Firstly,a hybrid parallel simulation engine architecture is proposed.Engine is the core component to support simulation execution.It is necessary to take advantages such as abundant threads,local shared memory and collabarative computing of CMP+MIC computing platforms for engines to gain better performances.At present,related works of multi-process ones are weak as they are in heavy granularity parallellism,large memory consumption and communication-consuming.Works on multi-core CPU platforms are hardly support simulation execution on multi-node platforms,while works aiming for other many core processors such as GPU are not compatible since the differences in architecture.To solve this problem,an engine architecture based on hybrid process and thread model(SE-HPTM)is proposed.SE-HPTM handles events in parallel inside both of CMP and MIC processes,while achieving collaborative simulation with inter processes communication such as MPI.A lock-free and memory-optimizing method is carried out to further improve the performance.Experimental results show that SE-HPTM brings 1.57 x speedup in compare with the multi-process engine architecture.Secondly,a computing model acceleration method based on gather-offload is proposed.Calculation processes of many complex models in simulation applications are ususlly time-consuming,which becomes the performance bottleneck of the whole system.However,the diversity and encapsulation of models makes it impossible to accelerate them with a common parallel or vecterization way,and the discrete scheduling events makes it hard to accelerate the mass processes in parallel.To solve this problem,a gather-offload model acceleration(GOMA)is proposed.GOMA is applied under the conservative execution of parallel simulations,aiming for the homogeneous models in parallel simulation.A subset of “safety events” will be gathered to aggregate the calculation process,and then offload them to MIC to calculate in parallel and vectorized.Experimental results show that GOMA brings as much as 3.4x speedup for a classical benchmark.Thirdly,a load balancing method is proposed based on Work-Stealing and objects regrouping.The load imbalance of parallel simulations usually becomes the performance bottleneck of the whole system.However,related works based on static distribution are not able to distribute load normally among processing units,while frequent dynamic migration will cause additional load into system.The global scheduling of events can achieve automatically load balancing while it is only for the single node multi-core platforms.To solve these problems,a load balancing method based on Work-Stealing and objects regrouping(WSReG)is proposed.WSReG is based on SE-HPTM,supporting collaborative simulation and not rely on any static division methods.Inside the process,WSReG uses the Work-Stealing way to balance load among threads,which reduce the migration of objects.Migration among processes only occurs when it is very necessary.Experimental results show that WSReG can effectively improve the system of load balancing degree,so as to improve the simulation application performance.Finally,a CMP+MIC based simulation environment(ECMIC)is proposed on the basis of the above works.A social opinion tendency simulation is designed to testify ECMIC.Experimental results show that ECMIC achieves almost 42% speedup in compare with the multi-process execution on CMP only.