Speaker
Description
In recent years, compute performances of GPUs (Graphics Processing Units) dramatically increased, especially in comparison to those of CPUs (Central Processing Units). GPUs are the hardware of choice for scientific applications involving massive parallel operations, such as deep learning (DL) and Artificial Intelligence (AI) workflows. High performance GPUs are offered by computing infrastructures owned by INFN or available to INFN researchers, such as on-premises data centers, HPC (High Performance Computing) centers, and public or private clouds. The programming paradigms for GPUs significantly vary according to the GPU model and vendor, often posing a barrier to their use in scientific applications. In addition, powerful GPUs, such as those available in HPCs or data centers, are hardly saturated by typical computing applications. Multiple vendors proposed custom solutions to allow for GPU partitioning. OpenForBC (Open For Better Computing) is an INFN-founded project (through the R4I 2020 call) with the aim of easing the use of such partitioning patterns. OpenForBC is an open source software framework that allows for effortless and unified partitioning of GPUs from different vendors in Linux KVM virtualized environments. OpenForBC supports dynamic partitioning for various configurations of the GPU, which can be used to optimize the utilization of GPU kernels from different users or different applications. For example training complex DL models may require a full GPU, but inference may need only a fraction of it, leaving free resources for multiple cloned instances or other tasks. In this contribution we describe the most common GPU partitioning options available on the market, discuss the implementation of the OpenForBC interface, and show the results of benchmark tests in typical use case scenarios.
