GTC 2020: Rebalancing the Load: Profile-Guided Optimization of the NAMD Molecular Dynamics Program for Modern GPUs using Nsight Systems
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Rebalancing the Load: Profile-Guided Optimization of the NAMD Molecular Dynamics Program for Modern GPUs using Nsight Systems
Julio Maia, University Of Illinois at Urbana-Champaign | David Hardy, University of Illinois at Urbana-Champaign
We'll show how we more than doubled the performance of the Nanoscale Molecular Dynamics program on modern GPUs through the use of profile-guided optimization with Nsight Systems and Nsight Compute. NAMD has historically offloaded most computation to GPUs, leaving a much smaller amount of work and device management to the CPU. However, the tremendous performance available on modern GPUs has led to an imbalance, in which the CPU work becomes a bottleneck while the GPU is left idling. We'll show how we improve GPU utilization and overall application performance by moving the remaining workload to the GPU while reducing CPU overhead due to communication and synchronization, all discovered with the help of Nsight Systems. We'll also discuss ongoing efforts in tuning the performance of our most compute-intensive kernels with Nsight Compute.