为提高分子动力学模拟中短程力的计算效率,设计并实现了基于FPGA的分子动力学模拟短程力多流水计算系统。针对在短程力多流水计算过程中多个计算模块频繁调用大量的粒子信息导致的高带宽需求和访问内存冲突问题,提出了多流水数据预取系统的设计,可减少对粒子数据的重复读取,缓解访问冲突,保证计算模块的效率。本文使用Xilinx Virtex UltraScale+HBM VCU128 FPGA开发板,实验结果表明,与短程力单流水计算系统相比,短程力多流水计算系统的计算效率提高了3.29倍,同时验证了多流水数据预取系统的有效性。
Abstract
To improve the computational efficiency of range-limited forces in molecular dynamics simulations,a range-limited force multi-pipeline computing system for molecular dynamics simulations based on FPGA is designed and implemented.To address the problems of high bandwidth requirements and access memory conflicts caused by the frequent calls of a large amount of particle information by multiple computational modules during the multi-pipeline computation of range-limited forces,the design of a multi-pipeline data prefetching system is proposed,which can reduce the repeated readings of particle data,alleviate access conflicts,and ensure the efficiency of computational modules.This paper uses Xilinx Virtex UltraScale+HBM VCU128 FPGA development board.The experiment results show that the computational efficiency of the multi-pipeline computing system is improved by 3.29 times compared with the single-pipeline computing system for range-limited force,and the effectiveness of the multi-pipeline data prefetching system is verified.
关键词
FPGA /
并行计算 /
短程力多流水计算系统 /
多流水数据预取系统
Key words
FPGA /
parallel computing /
range-limited force multi-pipeline computing system /
multi-pipeline data prefetching system
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] Xie T,Lanord A,Wang Y,et al.Graph dynamical networks for unsupervised learning of atomic scale dynamics in materials[J].Nature communications,2019,10(1):1-9.
[2] 王娟娟,李海平.分子模拟技术在食品分子互作中的应用研究进展[J].食品与发酵工业,2022,48(14):292-302.
[3] Poghosyan A,Astsatryan H,Wahi N Y.On the performance and energy consumption of molecular dynamics applications for heterogeneous CPU-GPU platforms based on Gromacs[J].Cybernetics and Information Technologies,2017,17(5):68-80.
[4] Chunshu W,Sahan B,Tong G.System level modeling of GPU/FPGA clusters for molecular dynamics simulations[C]//2021 IEEE High Performance Computing Conference,2021:1-8.
[5] Yang C,Geng T,Wang T,et al.Molecular dynamics range-limited force evaluation optimized for FPGAs[C]//2019 IEEE 30th International Conference on Application-Specific System,Arichitectures and Processors(ASAP),2019:263-271.
[6] Wu C,Geng T,Yang C,et al.A communication-efficient multi-chip design for range-limited molecular dynamics[C]//2020 IEEE High Performance Extreme Computing Conference,2020:1-8.
[7] Wu C,Geng T,Bandara S,et al.Upgrade of FPGA range-limited molecular dynamics to handle hundreds of processors[C]//2021 IEEE 29th Annual International Symposium on Field-Programmable Custom Computing Machines,2021:142-151.
[8] Chiu M,Khan M A,Herbordt M C.Efficient calculation of pairwise nonbonded forces[C]//2011 IEEE 19th Annual International Symposium on Field-Programmable Custom Computing Machines,2011:73-76.
[9] Yang C,Geng T,Wang T,et al.Fully integrated FPGA molecular dynamics simulations[C]//Proceedings of the International Conference for High Performance Computing,Networking,Storage and Analysis,2019:1-31.
基金
*高等学校学科创新引智计划项目(B12018);未来网络科研基金项目(FNSRFP2021YB11)。
PDF(1138 KB)
