Cloud computing, whether it concerns -geographically -distributed environments (grid computing) or computing centers with computer arrays (clusters), lies inevitably in the center of interest nowadays. For any application that requires increased computing power, efficiency of the above technologies is of particular importance (HPC). Efficient parallelization of repetitive tasks offered by the use of GPUs -instead of CPUs -(GPGPU) in computer clusters/grid computing for accelerating compute

more »

... sive applications, make it an increasingly adopted solution. The combination of GPU performance with the advantages offered by virtual environments (scalability, elasticity, security, live migration) offers both high performance and flexibility. At the same time it introduces problems that have thus far prevented wide adoption of such solutions despite them gaining great attention in research. Most importantly the performance overhead added by virtualization and the efficiency provided by a combination of that kind: one such solution should not only offer high performance, but also be economical (ie. offering the ability to use a GPU only in certain physical nodes of a computer cluster). The purpose of this thesis is to study the structure and ways to efficiently use such combinatorial systems that leverage GPGPU and virtualization. In particular, we focus on the ability of GPU-less computers to effectively run CUDA applications on remote (virtual or physical) GPU-equiped computers, namely simultaneous (efficient) use of a host GPU by multiple (remote) clients. In this context, we present an implementation (GPUsockets) that is "transparent"/without the need of any user-side modification, based on the split-driver model, that uses CUDA driver API, so that it offers ease of use and extensibility. We evaluate its' efficiency in a Xen environment and, to the extent that the results are satisfying, we propose further improvements (ie. by using alternative solution for inter-VM communication like V4V/V4VSockets) and possible future extensions.