Implementation And Optimization Of Molecular Dynamics Application On Sunway Taihulight Supercomputer

The Sunway TaihuLight is a new generation supercomputer by our own R&D,which has 125 peak petaflops across the entire machine with a LINPACK benchmark rating of 93 petaflops,and both performances are ranked number one in the world.The system uses 40960 SW26010 manycore processors with a total of 10,400,000 computing cores.The instruction set of Sunway processor is specified as ShenWei-64 with the autonomous core technology,and the processor runs at a clock speed of 1.45GHz.The peak floating point performance of each node reaches to 3.06Tflop/s.Through the construction of its own hardware and software system,the Sunway TaihuLight realizes the fully autonomous security controllable from software and hardware two aspects.However,the problem is that the existing application software cannot be operated on the system through simple transplantation,but it needs to be realized and optimized in combination with the structure characteristics of the system.Molecular Dynamics Software NAMD is an open source molecular dynamics simulation software based on the Charm++ parallel programming model,which is able to quickly simulate the massively parallel computers of millions of atomic macromolecules,and has won Gordon Bell Awards--one of the three major projects in the United State.NAMD is widely used in physics,chemistry,biology,materials,medicine and other fields.The aim of this paper is to analyze the architecture of NAMD software,the Charm++ parallel programming model and its runtime,then transplanting Charm++ on the Sunway TaihuLight and reimplement NAMD.Based on the structure characteristic of the Sunway TaihuLight and the ShenWei processor,the NAMD software is optimized both on the hardware and software,in order to fully exert the structure superiority of the system and improve the efficiency of parallel computation.Research and innovation of this paper includes the following three aspects:(1)Implement NAMD on the Sunway TaihuLight.Through taking a deep analysis of the characteristics of NAMD,such as software structure,hotspot,computation,memory access,communications and I/O,as well as the Charm++ parallel programming model and runtime,computing tasks are cut into slices with eliminating the dependency relations,thus achieving the scheme of implementing the NAMD software on ShenWei processor with parallelism.When migrating the third-party libraries which NAMD depends on,cross-compiled is used to implement the migration of TCL,FFTW and Charm++ on ShenWei processor.(2)Chip-level optimizations of NAMD on ShenWei processor with manycore architecture.Based on NAMD software runtime and the architecture of ShenWei processor,we propose an optimization scheme on single node.The NAMD software is optimized combining with hardware,including optimization of memory access,work modes inside the processors,and specified optimization based on the processors hardware features.Experiment has shown that after optimization,the performance of the program has gained about 20x.(3)System-level optimizations of NAMD on the Sunway TaihuLight.Targeting at the large scale systems structure of the Sunway TaihuLight,we propose a new task partition and load balance control model,which can solve the parallel efficiency problem under the large scale multi-machine.Through experiments,we got the result that under the new proposed model,when the computing scale is expanded to 5 million atoms,50,000 CGs(3.25 billion cores)can be utilized,and the program performance has been nearly doubled.