Abstract-Over the last few years, running high performance computing applications in the cloud has become feasible. At the same time, GPGPUs are delivering unprecedented performance for HPC applications. Cloud providers thus face the challenge to integrate GPGPUs into their virtualized platforms, which has proven difficult for current virtualization stacks.In this paper, we present LoGV, an approach to virtualize GPGPUs by leveraging protection mechanisms already present in modern hardware. LoGV enables sharing of GPGPUs between VMs as well as VM migration without modifying the host driver or the guest's CUDA runtime. LoGV allocates resources securely in the hypervisor which then grants applications direct access to these resources, relying on GPGPU hardware features to guarantee mutual protection between applications. Experiments with our prototype have shown an overhead of less than 4% compared to native execution.

In this paper, we present a light-weight, micro-kernel-based virtual machine monitor (VMM) for the Blue Gene/P Supercomputer. Our VMM comprises a small µ-kernel with virtualization capabilities and, atop, a user-level VMM component that manages virtual BG/P cores, memory, and interconnects; we also support running native applications directly atop the µ-kernel. Our design goal is to enable compatibility to standard OSes such as Linux on BG/P via virtualization, but to also keep the amount of kernel functionality small enough to facilitate shortening the path to applications and lowering OS noise.Our prototype implementation successfully virtualizes a BG/P version of Linux with support for Ethernet-based communication mapped onto BG/P's collective and torus network devices. First experiences and experiments show that our VMM still shows a substantial performance hit; nevertheless, our approach poses an interesting OS alternative for Supercomputers, providing the convenience of a fully-featured commodity software stack, while also promising to deliver the scalability and low latency of an HPC OS.

Over the last few years, GPUs have been finding their way into cloud computing platforms, allowing users to benefit from the performance of GPUs at low cost. However, a large portion of the cloud's cost advantage traditionally stems from oversubscription: Cloud providers rent out more resources to their customers than are actually available, expecting that the customers will not actually use all of the promised resources. For GPU memory, this oversubscription is difficult due to the lack of support for demand paging in current GPUs. Therefore, recent approaches to enabling oversubscription of GPU memory resort to software scheduling of GPU kernels -which has been shown to induce significant runtime overhead in applications even if sufficient GPU memory is available -to ensure that data is present on the GPU when referenced.In this paper, we present GPUswap, a novel approach to enabling oversubscription of GPU memory that does not rely on software scheduling of GPU kernels. GPUswap uses the GPU's ability to access system RAM directly to extend the GPU's own memory. To that end, GPUswap transparently relocates data from the GPU to system RAM in response to memory pressure. GPUswap ensures that all data is permanently accessible to the GPU and thus allows applications to submit commands to the GPU directly at any time, without the need for software scheduling. Experiments with our prototype implementation show that GPU applications can still execute even with only 20 MB of GPU memory available. In addition, while software scheduling suffers from permanent overhead even with sufficient GPU memory available, our approach executes GPU applications with native performance.

Over the last few years, GPUs have been finding their way into cloud computing platforms, allowing users to benefit from the performance of GPUs at low cost. However, a large portion of the cloud's cost advantage traditionally stems from oversubscription: Cloud providers rent out more resources to their customers than are actually available, expecting that the customers will not actually use all of the promised resources. For GPU memory, this oversubscription is difficult due to the lack of support for demand paging in current GPUs. Therefore, recent approaches to enabling oversubscription of GPU memory resort to software scheduling of GPU kernels -- which has been shown to induce significant runtime overhead in applications even if sufficient GPU memory is available -- to ensure that data is present on the GPU when referenced. In this paper, we present GPUswap, a novel approach to enabling oversubscription of GPU memory that does not rely on software scheduling of GPU kernels. GPUswap uses the GPU's ability to access system RAM directly to extend the GPU's own memory. To that end, GPUswap transparently relocates data from the GPU to system RAM in response to memory pressure. GPUswap ensures that all data is permanently accessible to the GPU and thus allows applications to submit commands to the GPU directly at any time, without the need for software scheduling. Experiments with our prototype implementation show that GPU applications can still execute even with only 20 MB of GPU memory available. In addition, while software scheduling suffers from permanent overhead even with sufficient GPU memory available, our approach executes GPU applications with native performance.

In this paper, we present a light-weight, micro-kernel-based virtual machine monitor (VMM) for the Blue Gene/P Supercomputer. Our VMM comprises a small µ-kernel with virtualization capabilities and, atop, a user-level VMM component that manages virtual BG/P cores, memory, and interconnects; we also support running native applications directly atop the µ-kernel. Our design goal is to enable compatibility to standard OSes such as Linux on BG/P via virtualization, but to also keep the amount of kernel functionality small enough to facilitate shortening the path to applications and lowering OS noise.Our prototype implementation successfully virtualizes a BG/P version of Linux with support for Ethernet-based communication mapped onto BG/P's collective and torus network devices. First experiences and experiments show that our VMM still shows a substantial performance hit; nevertheless, our approach poses an interesting OS alternative for Supercomputers, providing the convenience of a fully-featured commodity software stack, while also promising to deliver the scalability and low latency of an HPC OS.