Microsoft Technology Licensing, LLC

Apple Inc.

David S Bakin

Mukund Sankaranarayan

Jai Srinivasan

Jayachandran Raja

Sean C McDowell

William G Parry

Dale Alan Sather

Michael B Jones

Forrest Curtis Foltz

Terje Kristian Backman

Andy R Raffman

George H Shaw

Robin C B Speed

Robin C Speed

Robin Speed

Mark H Lucovsky

Mohammad Rastegari

Marc Shepard

David Michael Sauntry

Hessam Bagherinezhad

Upender Sandadi

Anish Jnyaneshwar Prabhu

Saman Naderiparizi

Andrew L Nicholson

Sayyed Karen Khatamifard

Sophie Lebrecht

Andrew Luke Chronister

James C Gabriel

Jonathan Gelsey

Andrew Z Luo

Systems and processes for training and compressing a convolutional neural network model include the use of quantization and layer fusion. Quantized training data is passed through a convolutional layer of a neural network model to generate convolutional results during a first iteration of training the neural network model. The convolutional results are passed through a batch normalization layer of the neural network model to update normalization parameters of the batch normalization layer. The convolutional layer is fused with the batch normalization layer to generate a first fused layer and the fused parameters of the fused layer are quantized. The quantized training data is passed through the fused layer using the quantized fused parameters to generate output data, which may be quantized for a subsequent layer in the training iteration.

Compressed convolutional neural network models

Resource management architectures implemented in computer systems to manage resources are described. In one embodiment, a general architecture includes a resource manager and multiple resource providers that support one or more resource consumers such as a system component or application. Each provider is associated with a resource and acts as the manager for the resource when interfacing with the resource manager. The resource manager arbitrates access to the resources provided by the resource providers on behalf of the consumers. A policy manager sets various policies that are used by the resource manager to allocate resources. One policy is a priority-based policy that distinguishes among which applications and/or users have priority over others to use the resources. A resource consumer creates an 'activity' at the resource manager and builds one or more 'configurations' that describe various sets of preferred resources required to perform the activity. Each resource consumer can specify one or more configurations for each activity. If multiple configurations are specified, the resource consumer can rank them according to preference. This allows the resource consumers to be dynamically changed from one configuration to another as operating conditions change.

Resource manager architecture utilizing a policy manager

Resource management architectures implemented in computer systems to manage resources are described. In one embodiment, a general architecture includes a resource manager and multiple resource providers that support one or more resource consumers such as a system component or application. Each provider is associated with a resource and acts as the manager for the resource when interfacing with the resource manager. The resource manager arbitrates access to the resources provided by the resource providers on behalf of the consumers, e.g., using a priority-based policy. A resource consumer creates an 'activity' at the resource manager and builds one or more 'configurations' that describe various sets of preferred resources required to perform the activity. Each resource consumer can specify one or more configurations, which may be ranked, for each activity. This allows the resource consumers to be dynamically changed from one configuration to another as operating conditions change.

Resource manager architecture

Methods, functional components and structures are disclosed for carrying out management of client/server processes operating within separate process spaces within a computer system. The disclosed methods, functional components facilitate and carry out management of client and server processes within a computer system such that a greater degree of control can be exhibited over their execution. An active server component process maintains a list structure identifying each client process that is currently using one of the server component process's interfaces. When a new client references a server, the list structure is augmented to include the system identification (e.g., handle) assigned by the operating system to the client process. A thread within the server process blocks while awaiting a triggering event (e.g., a client process in the list terminates). When the client process terminates, the block on the thread releases and the thread processes the change in client references. If the list of clients within the structure reaches zero, then the server process can immediately terminate.

Method and system for effective management of client and server processes

Resource management architectures implemented in computer systems to manage resources are described. In one embodiment, a general architecture includes a resource manager and multiple resource providers that support one or more resource consumers such as a system component or application. Each provider is associated with a resource and acts as the manager for the resource when interfacing with the resource manager. The resource manager arbitrates access to the resources provided by the resource providers on behalf of the consumers. A policy manager sets various policies that are used by the resource manager to allocate resources. One policy is a priority-based policy that distinguishes among which applications and/or users have priority over others to use the resources. A resource consumer creates an &ldquo;activity&rdquo; at the resource manager and builds one or more &ldquo;configurations&rdquo; that describe various sets of preferred resources required to perform the activity. Each resource consumer can specify one or more configurations for each activity. If multiple configurations are specified, the resource consumer can rank them according to preference. This allows the resource consumers to be dynamically changed from one configuration to another as operating conditions change.

A system is configured to prioritize streaming disk I/O over non-streaming disk I/O by providing high priority queuing to streaming disk I/O and/or to throttle non-streaming disk I/O when the total disk I/O (streaming&plus;non-streaming) exceeds a threshold amount for a given time quantum. When disk throttling is utilized, streaming disk I/O is processed in a first time quantum. Non-streaming disk I/O is processed, as much as possible, in the remainder of the first time quantum. Other non-streaming disk I/O remaining to be processed is deferred to a subsequent time quantum.

Disk throttling and priority queuing for streaming I/O

Unenforced allocation of disk and CPU bandwidth for streaming I/O

Growing community of inventors

Seattle, WA, United States of America

David S Bakin