OPEN FORBC

The project wants to develop an open source technology in a Linux environment for GPU partitioning.

Enabling these technologies in a Linux KVM environment would smooth out one of the most important differences that still play to the detriment of its adoption in professional and scientific contexts that need graphics computing power. This solution is aimed at both GPU computing applications and VDI (Virtual Desktop Infrastructure) applications. For both of them, the motivation for this drive, which is at the heart of the project, is to be able to improve on their currently adopted solutions in order to maximise their effectiveness, both economically and energetically.

From a computational point of view, GPU software very often requires a huge amount of work to be able to fully use the hardware available, which translates into huge investments for the development and optimization of algorithms, to allow their scalability. These developments require high specificity of the algorithms that often lead to a loss of software portability. However, if this goal is not pursued, the GPU is only partially used, wasting a significant fraction of peak performance and initial investment. The possibility of partitioning a GPU, through software definition, allows this limitation to be overcome, making it possible to use less optimized software by running multiple instances in parallel. With this approach, we expect to see better saturation of GPU power, which translates into higher hardware power efficiency and greater compatibility.

The same approach is applicable for the virtualisation of front-end machines, i.e. those whose operation takes place through screen-keyboard-mouse interaction. In this context, the frontier offered by this project is the possibility of increasing the number of users per server and the consequent improvement of energy and economic performance.

This solution can be adopted as a front-end for an HPC (High Performance Computing) cluster, allowing users to rely on a remote session equipped with graphics acceleration to display the results of calculations performed by the cluster.