Research On Parallel Technology For GROMACS Of Molecular Dynamics Simulation Software Based On CPU/MIC Cooperation On Tianhe-2 Supercomputer

Molecular Dynamics(MD)simulation is a computer simulation of physical movements of atoms and molecules to obtain the macroscopic properties of the atomic system by using Newton's classical mechanics.GROMACS is one of the most popular software for molecular dynamics simulation,which was developed by University of Groningen,and has been accelerated by several dozen versions of the optimization,including common MPI,GPU,OPENMP and SIMD,etc.GROMACS is powerful,user-friendly,and can meet the needs of almost all of the common atomic system simulation.At present,the problem of using GROMACS to perform molecular dynamics simulation is that the software running cycle is too long.In 2014,the Tianhe-2 supercomputer,which was developed by the National University of Defense Technology,was rated as the world's first supercomputer by Top500 with the performance of 33.86 petaflop/s.The Tianhe-2 supercomputer,which has 32000 CPU and 48000 MIC cards,is the latest super computing platform in the domestic field.1.On Tianhe-2 supercomputer,the GROMACS can only use CPU,which does not make full use of the computing resources.Aiming at this problem,we analyzed the problems and difficulties in the implementation of GROMACS with offload mode.Then,based on this,we speeded up the software by using data integration,multi computing unit,using new data stream,and finally finished the optimization of GROMACS with offload mode.This method can use the MIC on Tianhe-2 supercomputer to accelerate the GROMACS.The final test found that the optimized software was accelerated.2.In view of the low efficiency of GROMACS with offload mode,we found that the main reason is that the extra time cost is too high.By using the new data flow optimization,the method of multi-computing unit acceleration,we further reduced the additional overhead,improved the parallelism of the software,and let CPU and MIC cooperate to calculate the core computing module,realizing the parallel speedup of GROMACS with offload mode.3.From the practicality and the development of the GROMACS software,we speeded up the multi-node GROMACS.In the process of implementation,we offloaded the the local and non local potential to MIC,divided the data into four kinds of communication mode according to the different characteristics of data,reducing the communication cost,and finally realize the parallel acceleration of multi-node GROMACS with offload mode.In short,we first implemented the GROMACS with offload mode on Tianhe-2 supercomputer in the world,and achieved a good acceleration.