Over the last few years, virtualization has become an essential tool for
increasing security and dependability in computing systems. At the same
time, GPUs have been finding their way into general purpose computing,
where they can deliver unprecedented performance for certain classes of
applications. However, present-day GPUs have proven hard to integrate
into virtualized environments. Current virtualization stacks supporting
GPUs either require exclusive access to the GPU or introduce
considerable processing overhead -- up to 21~\% in case of the best
previous solution (vCUDA).

In my talk, I will present LoGV, a novel approach to virtualizing GPUs.
LoGV enables secure sharing of GPUs between VMs as well as VM migration
by leveraging protection mechanisms already present in modern GPU
hardware. LoGV allocates resources securely in the hypervisor, which
then grants guest VMs direct access to these resources, relying on GPU
hardware features to guarantee mutual protection between VMs.
Experiments with our prototype have shown an overhead of less than 4%
compared to native execution.