Data Management for Molecule Simulation : A Throughput-Oriented Approach

点击翻译按钮获取中文摘要

Project Summary This is an administrative supplement to the parent R01 project GM140316 titled “Data Management for Molecule Sim- ulation: A Throughput-Oriented Approach." The main objective is to acquire a parallel computing server that hosts 4 NVidia H100 GPU cards, two Intel Xeon Platinum 8480+ 28-core CPUs, 2TB of DDR5 system memory, and 64TB of NVMe M.2 internal storage. We believe this will be an essential component that leads to the success of the proposed research tasks in the parent grant. Specifically, the new server will provide the following capabilities that are not found in existing computing facilities the project team has access to: (1) and Memory Capacity: With 37-64 petaFLOPs of to- tal in-core computing power, 320GB of GPU memory with a bandwidth of 3.35TB/s, the server can significantly improve Specific Aim 2 tasks by reducing the time for processing analytical queries against molecular dynamics (MD) simula- tion data; (2) Storage Capacity: we propose to use NVMe devices with 64TB of space to store datasets and associated metadata, with 1.2 million IOPS and 48GBps of total bandwidth. This will allow low latency access to large datasets for processing, which is required by tasks in Specific Aim 3; (3) Data Communication: the H100 GPUs are connected by NVidia's 4th-generation NVLink buses, allowing non-blocking bidirectional communication between GPUs with a total bandwidth of 900GB/s. This will speed up large-scale MD simulations by releasing the GPU-to-GPU data transmission bottleneck, affecting Specific Aim 4 of the parent project.