Maximizing GPU utilization with NVIDIA Multi-Process-Service for Quantum Monte Carlo simulations with the TurboRVB code

This white paper describes the use of the NVIDIA Multi-Process Service (MPS) technology to enhance the performance of the Quantum Monte Carlo code TurboRVB on the Leonardo Booster partition at CINECA. The usage of NVIDIA MPS, allowing to efficiently schedule more than 1 MPI task per GPU, resulted in a factor-4 reduction in the GPU-hours needed to accomplish the simulations, without altering the wall time nor touching the source code. The reasons for this saving of resources are inspected with the NVIDIA profiling tools, and let us speculate that similar trends would be found in many scientific workloads other than Quantum Monte Carlo. This work was conducted as part of the EPICURE European initiative, providing high-level assistance to all the EuroHPC supercomputing projects.