Abstract: Techniques for right-sizing a GPU profile for a VM based on the VM's runtime behavior are provided. In one set of embodiments, these techniques can include collecting data regarding the VM's GPU resource usage and other performance/usage metrics, analyzing the collected data to predict the maximum amount of GPU memory and/or compute resources that the VM will likely require during its runtime, and determining a new, right-sized GPU profile for the VM based on the predicted maximum resource requirements.

Abstract: A framework that may be implemented by a workload orchestration platform for scheduling accelerator-enabled workloads on the accelerators in a cluster is provided. In one set of embodiments, the framework enables the platform to schedule accelerator-enabled workloads based on a multitude of user-provided, fine-grained accelerator requirements. In another set of embodiments, the framework enables the platform to automatically recommend an initial set of accelerator resource requirements for an accelerator-enabled workload and automatically right-size such requirements based on telemetry data collected during the workload's runtime.

Abstract: In one set of embodiments, a computer system can receive a plurality of requests for placing a plurality of clients on a plurality of graphics processing units (GPUs), where each request includes a profile specifying a number of GPU compute slices and a number of GPU memory slices requested by a corresponding client. The computer system can further formulate an integer linear programming (ILP) problem based on the requests and a maximum number of GPU compute and memory slices supported by each GPU. The computer system can then generate a solution for the ILP problem and place the plurality of clients on the plurality of GPUs in accordance with the solution.

Abstract: Disclosed are aspects of network function placement in virtual graphics processing unit (vGPU)-enabled environments. In one example a network function request is associated with a network function. A scheduler selects a vGPU-enabled GPU to handle the network function request. The vGPU-enabled GPU is selected in consideration of a network function memory requirement or a network function IO requirement. The network function request is processed using an instance of the network function within a virtual machine that is executed using the selected vGPU-enabled GPU.

Abstract: Disclosed are aspects of workload selection and placement in systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. In some aspects, workloads are assigned to virtual graphics processing unit (vGPU)-enabled graphics processing units (GPUs). A number of vGPU placement neural networks are trained to maximize a composite efficiency metric based on workload data and GPU data for the plurality of vGPU placement models. A combined neural network selector is generated using the vGPU placement neural networks, and utilized to assign a workload to a vGPU-enabled GPU.

Abstract: Various aspects are disclosed for graphics processing unit (GPU)-remoting latency aware migration. In some aspects, a host executes a GPU-remoting client that includes a GPU workload. GPU-remoting latencies are identified for hosts of a cluster. A destination host is identified based on having a lower GPU-remoting latency than the host currently executing the GPU-remoting client. The GPU-remoting client is migrated from its current host to the destination host.

Abstract: Disclosed are various embodiments for rate proportional scheduling to reduce packet loss in virtualized network function chains. A congestion monitor executed by a first virtual machine executed by a host computing device can detect congestion in a receive queue associated with a first virtualized network function implemented by a first virtual machine. The congestion monitor can send a pause signal to a rate controller executed by a second virtual machine executed by the host computing device. The rate controller can receive the pause signal. In response, the rate controller can pause the processing of packets by a second virtualized network function implemented by the second virtual machine to reduce congestion in the receive queue of the first virtualized network function.

Abstract: Exemplary methods, apparatuses, and systems include a client virtual machine processing a system call for a device driver to instruct a physical device to perform a function and transmitting the system call to an appliance virtual machine to execute the system call. The client virtual machine determines, in response to the system call, that an established connection with the appliance virtual machine has switched from a first protocol to a second protocol, the first and second protocols including a high-performance transmission protocol and Transmission Control Protocol and Internet Protocol (TCP/IP). The client virtual machine transmits the system call to the appliance virtual machine according to the second protocol. For example, the established connection may switch to the second protocol in response to the client virtual machine migrating to the first host device from a second host device.

Abstract: Disclosed are aspects of workload selection and placement in systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. In some aspects, workloads are assigned to virtual graphics processing unit (vGPU)-enabled graphics processing units (GPUs) based on a variety of vGPU placement models. A number of vGPU placement neural networks are trained to maximize a composite efficiency metric based on workload data and GPU data for the plurality of vGPU placement models. A combined neural network selector is generated using the vGPU placement neural networks, and utilized to assign a workload to a vGPU-enabled GPU.

Abstract: Disclosed are aspects of task assignment for systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. In some examples, an algorithm is determined based on predetermined virtual machine assignment algorithms. The algorithm optimizes for a predetermined cost function. A virtual machine is queued in an arrival queue for assignment. A graphics configuration of a system is determined. The graphics configuration specifies a number of graphics processing units (GPUs) in the system. The system includes a vGPU enabled GPU. The algorithm is selected based on a correlation between the algorithm and the graphics configuration of the system. The virtual machine is assigned to a run queue based on the selected algorithm.

Abstract: A method for symmetric authentication is provided. This method includes generating a first challenge message containing a first string; encrypting the first challenge message; transmitting the encrypted first challenge message to a second device; receiving a first answer message from the second device; decrypting the first answer message; and authenticating the second device based on determining the decrypted first answer message contains the first string. Upon successful authentication of the second device, the method further includes receiving an encrypted second challenge message from the second device; decrypting the encrypted second challenge message; generating a second answer message containing a second string; encrypting the second answer message; and transmitting the encrypted second answer message to the second device.

Abstract: A system and method for assessing video quality of a video-based application trains a neural network using training data of video samples and assesses video of the video-based application using the neural network to generate the subjective video quality information of the video-based application. Data augmentation is performed on video data, which is labeled with at least one subjective quality level, to generate the training data of video samples.

Abstract: Disclosed are aspects of virtual graphics processing unit (vGPU) scheduling-aware virtual machine migration. Graphics processing units (GPUs) that are compatible with a current virtual GPU (vGPU) profile for a virtual machine are identified. A scheduling policy matching order for a migration of the virtual machine is determined based on a current vGPU scheduling policy for the virtual machine. A destination GPU is selected based on a vGPU scheduling policy of the destination GPU being identified as a best available vGPU scheduling policy according to the scheduling policy matching order. The virtual machine is migrated to the destination GPU.

Abstract: Method and system for training a neural network. The neural network is split into first and second portions. A k-layer first portion is sent to a client training/inference engine and the second portion is retained by a server training/inference engine. At the splitting point, the kth layer is a one-way function in output computation has a number of nodes that are less than any other layer of the first portion. The client training/inference engine trains the first portion with input data in a set of training data. The server training/inference engine receives a batch of outputs from the client training and applies them to the second portion to train the entire neural network.

Abstract: Disclosed are aspects of memory-aware placement in systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. In some examples, graphics processing units (GPU) are identified in a computing environment. Graphics processing requests are received. A graphics processing request includes a GPU memory requirement. The graphics processing requests are processed using a graphics processing request placement model that minimizes a number of utilized GPUs that are utilized to accommodate the requests. Virtual GPUs (vGPUs) are created to accommodate the graphics processing requests according to the graphics processing request placement model. The utilized GPUs divide their GPU memories to provide a subset of the plurality of vGPUs.

Abstract: Disclosed are aspects of network function placement in virtual graphics processing unit (vGPU)-enabled environments. In one example a network function request is associated with a network function. A scheduler selects a vGPU-enabled GPU to handle the network function request. The vGPU-enabled GPU is selected in consideration of a network function memory requirement or a network function IO requirement. The network function request is processed using an instance of the network function within a virtual machine that is executed using the selected vGPU-enabled GPU.

Abstract: Disclosed are aspects of memory-aware placement in systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. Virtual graphics processing unit (vGPU) data is identified for graphics processing units (GPUs). A configured GPU list and an unconfigured GPU list are generated using the GPU data. The configured GPU list specifies configured vGPU profiles for configured GPUs. The unconfigured GPU list specifies a total GPU memory for unconfigured GPUs. A vGPU request is assigned to a vGPU of a GPU. The GPU is a first fit, from the configured GPU list or the unconfigured GPU list that satisfies a GPU memory requirement of the vGPU request.

Abstract: Various examples are disclosed for generating heatmaps and plotting utilization of hosts in a datacenter environment. A collector virtual machine can rove the datacenter and collect utilization data. The utilization data can be plotted on a heatmap to illustrate utilization hotspots in the datacenter environment.

Abstract: A system and method for assessing video quality of a video-based application inserts frame identifiers (IDs) into video content from the video-based application and recognizes the frame IDs from the video content using a text recognition neural network. Based on recognized frame IDs, a frame per second (FPS) metric of the video content is calculated. Based on the FPS metric of the video content, objective video quality of the video-based application is assessed.

Abstract: Various examples are disclosed for generating heatmaps and plotting utilization of hosts in a datacenter environment. A collector virtual machine can rove the datacenter and collect utilization data. The utilization data can be plotted on a heatmap to illustrate utilization hotspots in the datacenter environment.