Abstract: A computer-implemented method in a computing network of a number of processing nodes 1 to X, in the computing network neurons of a Convolutional Neural Network (CNN) are divided between the number of nodes. The method including allocating a mini-batch of input data from among mini-batches of input data to a node of the nodes; splitting the mini-batch into a number of mini-batch sections X corresponding and equal to the number of nodes; at the node retaining a mini-batch section which has a same number as the node and sending other mini-batch sections of the split mini-batch sections to corresponding other nodes according to a number of the split mini-batch sections; collating at the node the split mini-batch sections at the node into a single matrix and multiplying the collated matrix by the neurons to provide output data sections having one section of output data per each mini-batch.

Abstract: A computer program compiled for a machine learning accelerator hardware and associated with a default input data size is received. An execution of an operation of the computer program is initiated. It is identified that a data size of an input data of the operation is smaller than the default input data size. The smaller data size of the input data of the operation rather than the default input data size is caused to be transferred to the machine learning accelerator hardware for the input data of the operation.

Abstract: A parallel processing unit (PPU) can be divided into partitions. Each partition is configured to operate similarly to how the entire PPU operates. A given partition includes a subset of the computational and memory resources associated with the entire PPU. Software that executes on a CPU partitions the PPU for an admin user. A guest user is assigned to a partition and can perform processing tasks within that partition in isolation from any other guest users assigned to any other partitions. Because the PPU can be divided into isolated partitions, multiple CPU processes can efficiently utilize PPU resources.

Abstract: A parallel processing unit (PPU) can be divided into partitions. Each partition is configured to operate similarly to how the entire PPU operates. A given partition includes a subset of the computational and memory resources associated with the entire PPU. Software that executes on a CPU partitions the PPU for an admin user. A guest user is assigned to a partition and can perform processing tasks within that partition in isolation from any other guest users assigned to any other partitions. Because the PPU can be divided into isolated partitions, multiple CPU processes can efficiently utilize PPU resources.

Abstract: Image subunit based guest scheduling is disclosed. For example, a memory stores an image registry, which stores a plurality of reference entries each associated with subunits hosted on each node of a plurality of nodes. A scheduler executing on a processor manages deployment of guests to the plurality of nodes including a first node and a second node, where a first guest is associated with an image file that includes a first subunit and a second subunit. The image registry is queried for at least one node of the plurality of nodes hosting the first subunit and/or the second subunit and the first node is determined to host the first subunit. The first guest is scheduled to the first node based on the first node hosting the first subunit.

Abstract: An electronic apparatus includes a communication interface to communicate with external servers; and a processor to control the electronic apparatus by executing at least one instruction. The processor is configured to receive, from each of the external servers, values of bandwidths of a plurality of GPU pairs into which a plurality of GPUs included in the external servers are combined and information on whether the plurality of GPUs are used, based on a input job related to machine learning being received, identify a number of GPUs and a bandwidth value that are required for performing tasks included in the input job, and determine GPUs among the plurality of GPUs to perform the tasks based on the values of the bandwidths of the plurality of GPU pairs, the received information on whether the plurality of GPUs are used, and the number of GPUs and the bandwidth value.

Abstract: Techniques usable in optimization processing are described. A system includes an optimization processing unit (OPU). The OPU includes stochastic computing units and at least one programmable interconnect. Each of the stochastic computing units includes nodes and multiplication unit(s) configured to interconnect at least a portion of the nodes. The programmable interconnect(s) are configured to provide weights for and to selectably couple a portion of the stochastic computing units.

Abstract: Resegmenting chunks of data for load balancing is disclosed. A plurality of first chunks of data is received. The plurality of first chunks of data includes one or more entries that include raw data produced by a component of an information technology environment and that reflects activity in the information technology environment. The plurality of first chunks of data is resegmented into a plurality of second chunks of data based on a source type of the plurality of first chunks. A first subset of the plurality of second chunks of data is distributed to a first indexer of a set of indexers. An occurrence of a trigger event is determined, and in response to the trigger event, a second subset of the plurality of second chunks of data is distributed to a second indexer of the set of indexers.

Abstract: Embodiments of systems and methods for managing performance optimization of applications executed by an Information Handling System (IHS) are described. In an illustrative, non-limiting embodiment, an IHS may include computer-executable instructions to, for each of multiple resources used to execute a target application, receive one or more machine learning (ML)-based hints associated with each resource that have been generated by a ML-based optimization service, and generate one or more augmented hints for at least one of the resources using a ML-to-ML orchestration service. The ML-to-ML orchestration service then transmits the augmented hints to the ML-based optimization service that combines the augmented ML-based hints with the one or more internally generated hints to generate augmented profile recommendations that are, in turn, used to adjust one or more settings of the resource to optimize a performance of the target application executed by the resource.

Abstract: In general, embodiments are disclosed herein for tracking and allocating graphics hardware resources. In one embodiment, a software and/or firmware process constructs a cross-application command queue utilization table based on one or more specified command queue quality of service (QoS) settings, in order to track the target and current utilization rates of each command queue on the graphics hardware over a given frame and to load work onto the graphics hardware in accordance with the utilization table. Based on the constructed utilization table for a given frame, any command queues that have exceed their respective target utilization value may be moved to an “inactive” status for the duration of the current frame. For any command queues that remain in an “active” status for the current frame, work from those command queues may be loaded on to slots of the appropriate data masters of the graphics hardware in any desired order.

Abstract: A computer-implemented method for efficient and scalable enclave protection for machine learning (ML) programs includes tailoring at least one ML program to generate at least one tailored ML program for execution within at least one enclave, and executing the at least one tailored ML program within the at least one enclave.

Abstract: A system and method for providing dynamic device virtualization is herein disclosed. According to one embodiment, the computer-implemented method includes providing a hypervisor and one or more guest virtual machines (VMs). Each guest VM is disposed to run a guest user process and the hypervisor is split into a device hypervisor and a compute hypervisor. The computer-implemented method further includes providing an interface between the device hypervisor and the compute hypervisor. The compute hypervisor manages an efficient use of CPU and memory of a host and the device hypervisor manages a device connected to the host by exploiting hardware acceleration of the device.

Abstract: In some embodiments, a method includes: displaying, on a first client device, a plurality of tasks; identifying, by the first client device, a task from the plurality of tasks, the task transferrable to a second client device in communication with the first client device; and sending, by the first client device, metadata for the task to the second client device in response to input received by the first client device, the task including metadata to allowing the second client device to display the task in the same manner as the task was displayed by the first client device.

Abstract: The present disclosure is directed to methods and apparatus for evaluating resources that would be used by machine learning model(s) for purposes of implementing the machine learning model(s) on resource-constrained devices. For example, in one aspect, a plurality of layers in a machine learning model may be identified. A plurality of respective output sizes corresponding to the plurality of layers may be calculated. Based on the plurality of output sizes, a maximum amount of volatile memory used for application of the machine learning model may be estimated and compared to a volatile memory constraint of a resource-constrained computing device. Output indicative of a result of the comparing may be provided at one or more output components.

Abstract: A specialized in-memory database health check process is utilized to resolve dependencies in a resource indicating requirements for an instance of an in-memory database. Specifically, when an instance of an in-memory database is created in response to a request, a list of one or more component handlers are obtained. These component handlers are modular functions, separate from each other but potentially dependent on one or more other component handlers, and act to validate various requirements listed in a resource for the request. Each of the component handlers are executed individually during execution of a Reconcile function. To the extent that the execution of any component handlers in the list is unsuccessful, the Reconcile function is rerun for another iteration. These iterations continue until all component handlers report back as successful. Instance creation is then considered successful and the instance of the in-memory database can be utilized by users.

Abstract: Systems, methods, and apparatuses for resource monitoring identification reuse are described. In an embodiment, a system comprising a hardware processor core to execute instructions storage for a resource monitoring identification (RMID) recycling instructions to be executed by a hardware processor core, a logical processor to execute on the hardware processor core, the logical processor including associated storage for a RMID and state, are described.

Abstract: A scaling manager manages deques that track groups of preinitialized instances used to scale respective groups of active compute instances. Various techniques for deque management include a technique where a total instance quantity is preconfigured for the total number of instances assigned to both the group and the deque of preinitialized instances. As the size of the group grows for scale-ups, the size of the deque may go down. For example, the deque is not replenished when the group scales, but does expand when the group scales down. The total instance quantity may be bounded, in some examples, and an additional “buffer amount” of preinitialized instances may be implemented to provide a safety margin for burst scaling, which can be further enhanced by transferring instances between data structures of different groups of instances in some cases.

Abstract: Systems, methods, and interfaces for the management of virtual machine instances and other programmatically controlled networks are provided. The hosted virtual networks are configured in a manner such that a virtual machine manager of the virtual network may monitor activity such as user requests, network traffic, and the status and execution of various virtual machine instances to determine possible security assessments. Aspects of the virtual network may be assessed for vulnerabilities at varying levels of granularity and sophistication when a suspicious event or triggering activity is detected. Illustrative embodiments of the systems and methods may be implemented on a virtual network overlaid on one or more intermediate physical networks that are used as a substrate network.

Abstract: Methods, apparatus, and processor-readable storage media for determining capacity in storage systems using machine learning techniques are provided herein. An example computer-implemented method includes obtaining capacity-related data from a storage system; forecasting, for a given temporal period, capacity of one or more storage objects of the storage system by applying machine learning techniques to at least a portion of the capacity-related data; aggregating the forecasted capacity for at least portions of the one or more storage objects; determining, based on the aggregated forecasted capacity of the storage objects, whether at least a portion of the storage system will run out of capacity in connection with the given temporal period; and performing one or more automated actions based at least in part on the determination as to whether the at least a portion of the at least one storage system will run out of capacity.

Abstract: Methods, systems, and computer-readable media for resource usage restrictions in a time-series database are disclosed. Elements of a plurality of time series are stored into one or more storage tiers of a time-series database. The time series are associated with a plurality of clients of the time-series database. Execution of tasks is initiated using one or more resources of one or more hosts. The time-series elements represent inputs to the tasks. The tasks comprise a first task and a second task. A usage of the one or more resources by the first task is determined to violate one or more resource usage restrictions. Based at least in part on the usage, one or more actions are performed to modify the execution of the first task. The one or more actions increase an amount of the one or more resources available to the second task.