FPGA-based Parallelization Technology Of Molecular Dynamics Simulation

Molecular Dynamics(MD)simulation is a research method that uses the powerful computing power of computers to simulate the changes of the system composed of particles.However,with the expansion of the scale of the simulation system and the extension of the time scale,the computational time required for simulation increases exponentially due to the serial execution of CPU(Central Processing Unit,CPU)instructions and the architectural characteristics of shared memory.The current mainstream MD simulation software has gradually been unable to afford the huge overhead caused by the surge in computing time,which undoubtedly delays the progress of related basic research and application development.To this end,this paper transplants the relevant theories and algorithms of MD simulation to the Field Programmable Gate Array(FPGA)hardware platform executed in parallel.Analyze the requirements of the Range-Limited force that accounts for the largest proportion of computing time in the simulation process,and carry out parallel control design and implementation around its interaction with the data stored on the chip.For MD simulation algorithms,when performing Range-Limited force calculations,it is necessary to perform high-frequency iterations on the force data results,and the data must be reread each iteration.How to reduce the number of communication with the data storage and avoid the blockage of the channel,so as to improve the communication efficiency,is the key to speeding up the MD simulation process.First of all,this paper completes the design of data storage,and discusses the implementation of data fetching.In order to reduce the hardware system’s dependence on offchip devices and reduce the data communication overhead,the design concept that all associated data resides on the FPGA is first proposed.Then,based on the core algorithm of the Range-Limited force calculation part: NT(Neutral Territory)algorithm,the data requirement analysis of the Range-Limited force calculation part and the parallel task assignment of the auxiliary Range-Limited force calculation part are completed.Next,data structure design and data segmentation have been completed successively focus on data requirements and task allocation,and data parallelization has been realized.After that,two different schemes based on AXI4(Advanced extensible Interface 4)bus and BRAM(Block RAM)interface were proposed for data access.Secondly,focusing on the concept of task parallelization,this paper realizes the module parallelization design of the particle data reading and force data updating parts that interact frequently with the data storage in the short-range force calculation process.At the beginning,the design and implementation of the corresponding read and write control blocks have been completed according to the two memory access schemes,and the comparison and selection have been made in combination with the application scenarios.Afterwards,an asynchronous FIFO(First In First Out)is proposed to solve the problem of data loss when transmitting between modules.On this basis,the design and functional verification of the particle data reading and force data updating modules are completed.Finally,this paper builds an FPGA hardware acceleration platform for MD simulation,and builds a test framework for the MD simulation acceleration system.After that,the correctness of the results is verified,while achieving a better execution speed.This paper adopts the cooperative working mode of hardware as the main part and software as the auxiliary.The software is responsible for loading the initial data,and the Range-Limited force part of the molecular dynamic simulation,which consumes resources and time,is hardware accelerated on the FPGA platform.This is of great significance for promoting the realization of an autonomous and controllable large-scale molecular dynamics simulation software and hardware acceleration platform.