Federico Silla received the MS and Ph.D. degrees in Computer Engineering from Universitat Politècnica de València, Spain, in 1995 and 1999, respectively. He is currently a Full Professor at the Department of Computer Engineering at that university. He worked for two years at Intel Corporation. He is the coordinator of the rCUDA remote GPU virtualization project since it began in 2008. He has published more than 100 papers in conferences and journals, as well as 10 book chapters. He has been member of Program Committees in many of the most prestigious conferences in his area, including CCGRID, SC, PACT, ISCA, ICS, MICRO, etc. He is also an Associate Editor of JPDC journal since 2015.

Topic: Improving data center efficiency with GPU virtualization: the rCUDA case

Virtualization techniques have shown to report benefits to many computing facilities. In this regard, in addition to virtualizing an entire computer by using virtual machines, it is also possible to virtualize individual devices, such as a GPU. This mechanism is known as GPU virtualization.

Notice, however, that GPU virtualization can be further enriched if GPUs located in other cluster nodes are virtualized and remotely provided to applications requiring their services. In this way, it is possible to create pools of GPUs that can be dynamically assigned to the applications running in the cluster. This mechanism is referred to as remote GPU virtualization. The remote GPU virtualization technique provides many advantages to the use of GPUs. For instance, it allows to share a GPU among many applications without requiring the use of virtual machines in order to access the virtualized GPU. It is also possible to assign all the GPUs in the cluster to a single application running in one of the nodes. GPU migration is also possible, thus allowing to perform load balancing or even server consolidation. In summary, remote GPU virtualization makes the use of GPUs more flexible, at the same time that noticeable energy savings can be achieved.

This talk presents how to improve the efficiency of data centers by using the rCUDA middleware, which implements the remote GPU virtualization mechanism. The latest developments within the rCUDA technology will be introduced along with many performance and energy studies that demonstrate that using remote GPU virtualization provides many benefits.