Scalability Analysis of Molecular Dynamics Simulation Using NAMD on Ampere-Based Dense GPU Supercomputer

Nanoscale molecular dynamics (NAMD) is a widely used scalable scientific software for molecular dynamics applications that simulate the movements of atoms and molecules in a bio-molecular system with millions of atoms. The goal of this paper is to study the scalability of NAMD on a dense GPU-based supercomputer named PARAM Siddhi-AI. Each node of the supercomputer is built with the latest NVIDIA Ampere-based A100 GPUs and AMD 7742 CPUs with 128 CPU cores. The scalability study is classified into two parts, i.e., intra-node GPU scalability and inter-node GPU scalability. The performance is analyzed for two NAMD versions, i.e., 2.13 and 3.0 Alpha for eight different input datasets, however, version 3.0 Alpha only supports single-node GPU runs. The input datasets ApoA1 with 92 thousand atoms whereas STMV with 1.06 billion atoms are used to perform the benchmarking. It is observed that NAMD version 3.0 Alpha performed better than version 2.13 for intra-node GPU scalability. Version 3.0 Alpha is ~ 1.5x to ~ 2x times more efficient on a node with 8 GPUs for the input datasets ApoA1 and STMV respectively. In comparison with CPU performance, the NAMD version 3.0 Alpha shows a speed-up of 22x and 90x whereas version 2.13 shows a speed-up of 19x and 12x for both the input datasets ApoA1 and STMV respectively. In the case of inter-node GPU performance, version 2.13 shows the speed-up of 6.98x and 6x across 4 nodes with 32 GPUs (each node with 8 GPUs) in comparison with CPU performance for input datasets ApoA1 and STMV respectively. The difference in the performance of version 3.0 Alpha and 2.13 is due to the elimination of the CPU bottleneck in NAMD 3.0 Alpha, and this is achieved by offloading almost all (force computations, integrator, and rigid bond constraints) computations to the GPU. The scalability analysis/performance study is also performed for the latest NVIDIA Ampere-based A100 GPUs v/s previous-generation NVIDIA Volta-based V100 GPUs. NAMD is 1.3x and 1.6x times more efficient or faster on NVIDIA A100 GPU in comparison with NVIDIA V100 GPUs for both classes of input datasets ApoA1 and STMV respectively.