C. Pascoe, Lawrence C. Stewart, B. W. Sherman, Vipin Sachdeva, Martin C. Herbordt
{"title":"在多个fpga上执行完整的分子动力学模拟","authors":"C. Pascoe, Lawrence C. Stewart, B. W. Sherman, Vipin Sachdeva, Martin C. Herbordt","doi":"10.1109/HPEC43674.2020.9286155","DOIUrl":null,"url":null,"abstract":"We have modified the open source molecular dynamics (MD) simulation code OpenMM [1] to add support for running complete MD timesteps on a cluster of FPGAs. The overall structure of the application is shown in Figure 1. MD proceeds by calculating forces on individual particles and integrating those forces to update velocities/positions on a per timestep basis. A variety of forces apply to each particle and we subdivide them into three categories based on the computation requirements: range limited (RL), long range (LR), and bonded. RL interactions comprise Lennard Jones and electrostatic forces between all particle pairs within a radial cutoff. LR interactions comprise electrostatic forces beyond the RL cutoff, where pairwise computation would be too costly. We calculate LR forces using the Smooth Particle Mesh Ewald (PME) method, which uses 3D Fast Fourier Transforms (FFTs) to accelerate computation. Bonded interactions are the focus of future work. Kernels are coded in OpenCL for ease of hardware development and application integration. The design uses a mix of fixedpoint and single-/double-precision floating-point arithmetic where needed to maintain the same level of accuracy as CPU and GPU implementations. The ultimate goal of this project is to perform MD simulation of biologically-relevant systems within the context of drug discovery (i.e., periodic systems of 50,000–100,000 particles with approximate density of 1 atom per 10 cubic Å) with strong scaling performance greater than possible with other technologies such as GPUs.","PeriodicalId":168544,"journal":{"name":"2020 IEEE High Performance Extreme Computing Conference (HPEC)","volume":"28 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2020-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"5","resultStr":"{\"title\":\"Execution of Complete Molecular Dynamics Simulations on Multiple FPGAs\",\"authors\":\"C. Pascoe, Lawrence C. Stewart, B. W. Sherman, Vipin Sachdeva, Martin C. Herbordt\",\"doi\":\"10.1109/HPEC43674.2020.9286155\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"We have modified the open source molecular dynamics (MD) simulation code OpenMM [1] to add support for running complete MD timesteps on a cluster of FPGAs. The overall structure of the application is shown in Figure 1. MD proceeds by calculating forces on individual particles and integrating those forces to update velocities/positions on a per timestep basis. A variety of forces apply to each particle and we subdivide them into three categories based on the computation requirements: range limited (RL), long range (LR), and bonded. RL interactions comprise Lennard Jones and electrostatic forces between all particle pairs within a radial cutoff. LR interactions comprise electrostatic forces beyond the RL cutoff, where pairwise computation would be too costly. We calculate LR forces using the Smooth Particle Mesh Ewald (PME) method, which uses 3D Fast Fourier Transforms (FFTs) to accelerate computation. Bonded interactions are the focus of future work. Kernels are coded in OpenCL for ease of hardware development and application integration. The design uses a mix of fixedpoint and single-/double-precision floating-point arithmetic where needed to maintain the same level of accuracy as CPU and GPU implementations. The ultimate goal of this project is to perform MD simulation of biologically-relevant systems within the context of drug discovery (i.e., periodic systems of 50,000–100,000 particles with approximate density of 1 atom per 10 cubic Å) with strong scaling performance greater than possible with other technologies such as GPUs.\",\"PeriodicalId\":168544,\"journal\":{\"name\":\"2020 IEEE High Performance Extreme Computing Conference (HPEC)\",\"volume\":\"28 1\",\"pages\":\"0\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2020-09-22\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"5\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"2020 IEEE High Performance Extreme Computing Conference (HPEC)\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/HPEC43674.2020.9286155\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"2020 IEEE High Performance Extreme Computing Conference (HPEC)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/HPEC43674.2020.9286155","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Execution of Complete Molecular Dynamics Simulations on Multiple FPGAs
We have modified the open source molecular dynamics (MD) simulation code OpenMM [1] to add support for running complete MD timesteps on a cluster of FPGAs. The overall structure of the application is shown in Figure 1. MD proceeds by calculating forces on individual particles and integrating those forces to update velocities/positions on a per timestep basis. A variety of forces apply to each particle and we subdivide them into three categories based on the computation requirements: range limited (RL), long range (LR), and bonded. RL interactions comprise Lennard Jones and electrostatic forces between all particle pairs within a radial cutoff. LR interactions comprise electrostatic forces beyond the RL cutoff, where pairwise computation would be too costly. We calculate LR forces using the Smooth Particle Mesh Ewald (PME) method, which uses 3D Fast Fourier Transforms (FFTs) to accelerate computation. Bonded interactions are the focus of future work. Kernels are coded in OpenCL for ease of hardware development and application integration. The design uses a mix of fixedpoint and single-/double-precision floating-point arithmetic where needed to maintain the same level of accuracy as CPU and GPU implementations. The ultimate goal of this project is to perform MD simulation of biologically-relevant systems within the context of drug discovery (i.e., periodic systems of 50,000–100,000 particles with approximate density of 1 atom per 10 cubic Å) with strong scaling performance greater than possible with other technologies such as GPUs.