Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: Techniques for signature verification of field-programmable gate array (FPGA) programs are described herein. In one or more implementations, an FPGA virtualization manager of a host device receives a request from a virtual machine for an FPGA program to program FPGAs of the host. The FPGA program is configured to program the FPGAs to provide functionality of a hardware-implementation of a respective program (e.g., a machine-learning algorithm) or of a respective device (e.g., a graphics processing unit). Before allowing the FPGA program to program the FPGAs, however, the FPGA virtualization manager determines whether the FPGA program is trusted to do so. To do so, the FPGA virtualization manager verifies a digital signature associated with the FPGA program. When the signature is verified the FPGA program is determined to be trusted. Based on such a determination, the FPGA virtualization manager loads the FPGA program to program the FPGAs to provide the functionality.

Abstract: A computing device includes a host, one or more guest partitions, and one or more physical devices. A physical device can be virtualized, at least in part, by the host and made available to the guest partitions. A physical device includes both a control plane and a data plane. The host provides direct access to at least part of the data plane of a physical device to a guest partition. However, the host virtualizes the control plane of the physical device, exposing a control plane for the physical device to the guest partition that is not the actual control plane of the physical device. Requests to access (e.g., read, write, modify, etc.) the control plane of the physical device are received by the host from the guest partition, and converted as appropriate to the control plane for the physical device.

Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: A computing device includes a host, one or more guest partitions, and one or more physical devices. A physical device can be virtualized, at least in part, by the host and made available to the guest partitions. A physical device includes both a control plane and a data plane. The host provides direct access to at least part of the data plane of a physical device to a guest partition. However, the host virtualizes the control plane of the physical device, exposing a control plane for the physical device to the guest partition that is not the actual control plane of the physical device. Requests to access (e.g., read, write, modify, etc.) the control plane of the physical device are received by the host from the guest partition, and converted as appropriate to the control plane for the physical device.

Abstract: Techniques for graphics processing unit (GPU) partitioning for virtualization are described herein. In one or more implementations, a GPU partitioning manager of a host device obtains a request for a virtual machine having GPU functionality. In particular, the request specifies the GPU functionality in terms of different GPU capabilities. These different capabilities correspond to segments of a GPU model that represents GPU functionality and is used to govern interactions between virtual machines and GPUs. The GPU partitioning manager determines whether GPUs of the host device are available to satisfy the request based on the specified capabilities. If so, the GPU partitioning manager allocates a portion of the determined available GPUs to the virtual machine to configure the virtual machine with a GPU partition having the functionality. The virtual machine configured with the GPU partition can then be exposed to provide GPU-processed data to a GPU partition requestor.

Abstract: Techniques for signature verification of field-programmable gate array (FPGA) programs are described herein. In one or more implementations, an FPGA virtualization manager of a host device receives a request from a virtual machine for an FPGA program to program FPGAs of the host. The FPGA program is configured to program the FPGAs to provide functionality of a hardware-implementation of a respective program (e.g., a machine-learning algorithm) or of a respective device (e.g., a graphics processing unit). Before allowing the FPGA program to program the FPGAs, however, the FPGA virtualization manager determines whether the FPGA program is trusted to do so. To do so, the FPGA virtualization manager verifies a digital signature associated with the FPGA program. When the signature is verified the FPGA program is determined to be trusted. Based on such a determination, the FPGA virtualization manager loads the FPGA program to program the FPGAs to provide the functionality.