Microsoft Technology Licensing, LLC

Hewlett-packard Development Company, L.p.

Digital Equipment Corporation

Google Inc.

Compaq Computer Corporation, Inc.

Other

University of Washington

Chandramohan Amyangot Thekkath

Lidong Zhou

William Henry Gates, Iii

Henricus Johannes Maria Meijer

Alexander G Gounares

Lili Cheng

Raymond Edward Ozzie

Gary William Flake

James R Larus

Amit Mital

Debi P Mishra

Daniel S Glasser

Ira L Snyder, Jr

Nishant Dani

Arnold Neil Blinn

Matthew Bret MacLaurin

Christopher Wellington Brumme

Marc Alexander Najork

Thomas F Bergstraesser

John Philip MacCormick

Mark David Lillibridge

Michael J Connolly

Thomas Lee Rodeheffer

Melora Zaner-Godsey

Dane Alexander Glasgow

Jeffrey Adgate Dean

Ranveer Chandra

Sanjay Ghemawat

Yonghui Wu

Mark Steven Manasse

Ruoming Pang

Michael Isard

Paul Ronald Barham

Darrell Anderson

Marcus Jon Jager

David Rogers Treadwell, Iii

Alice Silverberg

Kristof Roomp

Sudip Roy

Ryan Sepassi

William Raymond Hoffman

Aakanksha Chowdhery

Laurent El Shafey

Brennan Saeta

Daniel William Hurt

Hyeontaek Lim

Parker Edward Schuh

Edward K Lee

Galen Clyde Hunt

Li Zhang

William J Bolosky

Kentaro Toyama

Daniel J Scales

Zheng Zhang

Edward Kihyen Lee

Roy Levin

Dahlia Malkhi

Yinglian Xie

Rohan Narayan Murty

Vijayan Prabhakaran

Raymond Paul Stata

Henry M Levy

Shiding Lin

Maya Haridasan

William Eugene Morgan

Venugopalan Ramasubramanian

Omid Fatemieh

Anna R Karlin

Timothy P Mann

Michael J Feeley

Nicholas Charles Murphy

Parya Moinzadeh

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for distributing machine learning workloads, e.g., computations for training a neural network or computing an inference using a neural network, across multiple hardware accelerators. One of the systems comprises a plurality of accelerator islands, each hardware accelerator island comprising a respective plurality of hardware devices that include a plurality of hardware accelerators and a corresponding host for each of the plurality of hardware accelerators; and a respective scheduler for each of the accelerator islands that is configured to schedule workloads across the plurality of accelerators and corresponding hosts in the accelerator island, wherein the system is configured to: receive data representing a machine learning workload; and assign a respective portion of the machine learning workload to each of the plurality of accelerator islands for scheduling by the respective scheduler for the accelerator island.

Asynchronous distributed data flow for machine learning workloads

Mobile devices periodically send resource usage reports to an energy server. The reports include information such as an amount of energy used during a time interval, and identify processes that are executed on the mobile devices during the time interval. These reports are analyzed by the energy server and are used to generate statistics regarding the energy used by the mobile devices and the processes that are executed by the mobile devices. When a mobile device experiences a suspected energy spike, the mobile device generates and sends a report to the energy server. After confirming the energy spike, the energy server uses the statistics and the received report to identify one or more causes of the energy spike. The causes can include one or more processes, or the mobile device itself. The mobile device can reduce the amount of energy used based on the identified causes.

Determining cause of energy spike using energy reports

The location of a user over time is monitored by a mobile device. The monitored locations are organized into tracks that describe a path or route that the user took over a period of time. Segments that correspond to each of the tracks are determined. The segments may correspond to road on a map, or some other standardization. The segments are associated with their corresponding tracks, and used to identify similar tracks or to generate similarity scores for pairs of tracks.

Comparing and identifying similar tracks

Remote provisioning of an IT network and/or associated services is provided. Hardware, software, service and/or expertise can be moved from on-premise to a remote location (e.g., central, distributed . . . ). Accordingly, at least a large degree computation can be moved to the center to exploit economies of scale, among other things. In such an architecture, computational resources (e.g., data storage, computation power, cache . . . ) can be pooled, and entities can subscribe to a particular level of resources related to a private entity IT network.

Remote provisioning of information technology

An enhancement to a web browser offers an extension mechanism for web applications to utilize client-side resources, thereby moving more extensibility and flexibility to the client-side. The web browser may expose some control interfaces to meet various requirements from web applications. Using the extension mechanism, the web applications are able to offload the cloud servers, function when offline, and improve performance. The extension mechanism also provides users with full control to the data passing through their web browsers.

Extensible browser platform for web applications

Personal data mining mechanisms and methods are employed to identify relevant information that otherwise would likely remain undiscovered. Users supply personal data that can be analyzed in conjunction with data associated with a plurality of other users to provide useful information that can improve business operations and/or quality of life. Personal data can be mined alone or in conjunction with third party data to identify correlations amongst the data and associated users. Applications or services can interact with such data and present it to users in a myriad of manners, for instance as notifications of opportunities.

Personal data mining

Multiple versions of data on different sets of machines allow a system to degrade gracefully even when experiencing excessive failures. When excessive failures cause the latest versions to be unavailable, the system becomes degraded, but still offers an old version if available. A most recent coherent set of the available older versions is found and provided. The degree of degradation increases gradually as the system experiences more and more failures. Graceful degradation is desirably complementary to fault tolerance.

Gracefully degradable versioned storage systems

An information management system includes a database comprising a B-Tree data structure. The B-Tree includes a plurality of nodes associated with disk blocks, handles stored in the nodes, and version numbers attached to each of the handles and disk blocks. In addition, a method is provided that can include initially assigning version numbers to disk blocks, wherein the disk blocks are associated with nodes in a B-Tree data structure, and then performing an allocate operation whereby a handle identifying a disk block is returned along with the disk block's version number.

Deletion and compaction using versioned nodes

Disclosed herein are scheduling techniques for transmitting time-critical data in a cost-aware manner over a network comprising a plurality of heterogeneous transmission interfaces. The scheduling problem is formulated as a linear programming problem with the deliver-by deadlines of the various data blocks as hard constraints and minimizing cost set as an objective (soft) constraint. The problem is simplified by assuming data blocks with the earliest deadlines should be scheduled first and the most aggressive interfaces should be used first. To formulate the linear programming problem, the time domain is divided into bins and various bin-level schedules are enumerated for switching the transmission of the data over various transmission interfaces. The linear programming techniques are applied to the various bin configurations and the least costly of the resulting transmission schedule is selected for submission to a switching layer.

Cost-aware networking over heterogeneous data channels

Methods and apparatuses are provided for recovering one or more replicated datasets. In accordance to some embodiments of the invention, this is accomplished by determining whether a log contains one or more entries indicating a corruption in a replicated data item; and recovering the replicated data item if the log contains any such entries. The log, however, is maintained at a higher layer in a software hierarchy than the layer at which the replicated data item is recovered. As a result, there is no need to maintain a separate replication log at the layer where the replicated data item is recovered. In accordance to some embodiments of the invention, this is accomplished by recording an entry indicating that an operation was performed on a replicated record in a log at a higher layer in a software hierarchy, determining whether a failure occurred during the operation, and reconciling the replicated record at a lower layer in the software hierarchy if it is determined that a failure occurred during the operation. As a result, the present invention eliminates the need for additional disk writes required by conventional techniques that use a replication log to at the same layer where the data recovery is performed.

Efficient recovery of replicated data items

Tree data storage structures are implemented on respective computers in a distributed environment, such as on a network, so that information associated with nodes of one computer's tree data storage structure may be read or written to by another computer in the network. To promote efficiency, a cache may be employed on the computers in the network such that each computer caches information associated with nodes of tree data storage structures located on the computers in the network. A lock service may implement a caching protocol to provide efficient concurrency of caching operations while ensuring that current information associated with the nodes is available to all computers in the network.

Implementing a tree data storage structure in a distributed environment

High throughput in data computations and processing is maintained while minimizing latency. A binary tree architecture is provided in which two trees are used simultaneously, and initiation of the trees is staggered to allow for optimal use of bandwidth. These techniques are desirable for erasure codes and other computations where the addition operator is commutative. Additionally, a ternary tree architecture may be used, in which three trees co-exist on the same set of nodes to maintain high throughput while further improving latency.

Erasure coding and group computations using rooted binary and ternary trees

The claimed subject matter provides systems and/or methods that facilitate replicating a state associated with a client, user, service, application, and the like. A third party service provider can support any number of services that can be concurrently requested by several clients without user perception of degraded computing performance as compared to conventional systems/techniques due to improved connectivity and mitigated latencies. A replication component can generate replicas of states associated with requested services. Further, the replicas can facilitate seamlessly interacting with the third party service provider (e.g., while transitioning between client devices). Additionally, by providing replicas of the state related information, differing third party service providers can effectuate services based upon a request from a client without regenerating the state.

State reflection

A searchable storage system indexes data, such as a blob (binary large object) as it is stored. The indexes are stored simultaneously with the data by the same storage mechanism. Data cannot be stored without indexing it. Any indexing technique may be used. The indexes are then used to access the data. This ensures that any inconsistencies can handled at the local level, and are not visible to clients of the system. Data can be put in to, or taken out from, the searchable storage system. Additionally, a query on the data in the searchable storage system may be performed.

Searchable storage system

Growing community of inventors

Palo Alto, CA, United States of America

Chandramohan Amyangot Thekkath