Google Inc.

Digital Equipment Corporation

Hewlett-packard Company

Compaq Computer Corporation, Inc.

Alta Vista Company

Sanjay Ghemawat

Jeffrey Adgate Dean

Yasushi Saito

Andrew Fikes

Howard Bradley Gobioff

Shun-Tak Leung

Gautham K Thambidorai

Benedict A Gomes

Sean Quinlan

Anurag Acharya

Sean Matthew Dorward

Daniel Wesley Dulitz

Deborah Anne Wallach

Robert C Pike

Bwolen Po-Jen Yang

Michael Burrows

Monika Hildegard Henzinger

Olcan Sercinoglu

Alexandre A Verstak

Fritz Schneider

Eric Russell Fredricksen

Ivan Posva

Huican Zhu

Srdjan Petrovic

Benjamin Thomas Smith

Fay Wen Chang

Matthew Rosencrantz

Simon Tong

Krishna Asur Bharat

Georges R Harik

Alexander Lloyd

Sergey Brin

Wilson Cheng-Yi Hsieh

Yuan Yu

Sachin Jain

Boris Mazniker

Vijay Vasudevan

Niels Provos

Paul A Tucker

Urs Hoelzle

Sebastian Kanthak

Craig Silverstein

John Abraham Bauer

Chris Jorgen Taylor

Wilson C Hsieh

Yonghui Wu

Mehran Sahami

Jay Michael Ponte

Ruoming Pang

Kourosh Gharachorloo

Michael Isard

Chandramohan Amyangot Thekkath

Michael James Boyer Epstein

Vibhu Mittal

John Piscitello

Paul Ronald Barham

Matt Cutts

David Konerding

Sudip Roy

Ryan Sepassi

Aakanksha Chowdhery

Laurent El Shafey

Michal Piotr Szymaniak

Jonathan Bingham

Gauthaum K Thambidorai

Christopher D Manning

Brennan Saeta

Daniel William Hurt

Hyeontaek Lim

Parker Edward Schuh

Christopher Cunningham Frost

Carl Alan Waldspurger

Richard Lee Sites

Ciprian Ioan Chelba

Andrei Z Broder

Martin Abadi

William Edward Weihl

Luiz Andre Barroso

Raymond Paul Stata

Keith H Randall

Boulos Harb

David L desJARDINS

Christopher Scott Taylor

Martin Farach-Colton

Lance M Berc

Mark Thierry Vandervoorde

Suresh Venkatasubramanian

Puneet Kumar

Yaz Saito

Georges R Hank

Gautham Thambidorai

Paul Tucker

Andrew B Fikes

John A Bauer

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

A method for deleting obsolete files from a file system is provided. The method includes receiving a request to delete a reference to a first target file of a plurality of target files stored in a file system, the first target file having a first target file name. A first reference file whose file name includes the first target file name is identified. The first reference file is deleted from the file system. The method further includes determining whether the file system includes at least one reference file, distinct from the first reference file, whose file name includes the first target file name. In accordance with a determination that the file system does not include the at least one reference file, the first target file is deleted from the file system.

Method and system for deleting obsolete files from a file system

A method of accessing data includes storing a table that includes a plurality of tablets corresponding to distinct non-overlapping table portions. Respective pluralities of tablet access objects and application objects are stored in a plurality of servers. A distinct application object and distinct tablet are associated with each tablet access object. Each application object corresponds to a distinct instantiation of an application associated with the table. The tablet access objects and associated application objects are redistributed among the servers in accordance with a first load-balancing criterion. A first request directed to a respective tablet is received from a client. In response, the tablet access object associated with the respective tablet is used to perform a data access operation on the respective tablet, and the application object associated with the respective tablet is used to perform an additional computational operation to produce a result to be returned to the client.

Associating application-specific methods with tables used for data storage

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for receiving a request from a client to process a computational graph; obtaining data representing the computational graph, the computational graph comprising a plurality of nodes and directed edges, wherein each node represents a respective operation, wherein each directed edge connects a respective first node to a respective second node that represents an operation that receives, as input, an output of an operation represented by the respective first node; identifying a plurality of available devices for performing the requested operation; partitioning the computational graph into a plurality of subgraphs, each subgraph comprising one or more nodes in the computational graph; and assigning, for each subgraph, the operations represented by the one or more nodes in the subgraph to a respective available device in the plurality of available devices for operation.

Processing computational graphs

A method performs large-scale data processing in a distributed and parallel processing environment. The method defines application-independent map and reduce operations, each invoking one or more library functions that automatically handle data partitioning, parallelization of computations, and fault tolerance. A user specifies a map operation, which calls one or more of the application-independent map operators to perform data read and write operations. A user also specifies a reduce operation, which calls one or more of the application-independent reduce operators to perform data read and write operations. The method executes application-independent map worker processes. Each map worker process executes the user-specified map operation to read designated portions of input files and store intermediate data values in intermediate data structures. The method also executes application-independent reduce worker processes. Each reduce worker process executes the user-specified reduce operation to read intermediate data values from the intermediate data structures and produce final output data.

System and method for large-scale data processing using an application-independent framework

A programming model provides a method for type inference in programming operations. Information defining one or more attributes of an operation is received, the information specifying a field including a field name and a field type identifier for each of the attributes. Constraints for the operation are determined at least based on the attributes, wherein the constraints restrict at least one of a type of input for the operation or a type of output for the operation. Information defining an input for the operation is received, and it is determined, based on the constraints and the received information defining the input, the type of output for the operation. The type of output is associated with an output for the operation.

System of type inference for tuple graph programs

Systems and methods for analyzing input data records are provided in which a master process initiates a plurality of concurrent first processes each of which comprises, for each data record in at least a subset of a plurality of input data records, creating a parsed representation of the data record and independently applying a procedural language query to the parsed representation to extract one or more values. A respective emit operator is applied to at least one of the extracted one or more values thereby adding corresponding information to a respective intermediate data structure. The respective emit operator implements one of a predefined set of statistical information processing functions. The master process also initiates a plurality of second processes each of which aggregates information from a corresponding subset of intermediate data structures to produce aggregated data that is, in turn, combined to produce output data.

System and method for analyzing data records

A system facilitates the distribution and redistribution of chunks of data among multiple servers. The system may identify servers to store a replica of the data based on at least one of utilization of the servers, prior data distribution involving the servers, and failure correlation properties associated with the servers, and place the replicas of the data at the identified servers. The system may also monitor total numbers of replicas of the chunks available in the system, identify chunks that have a total number of replicas below one or more chunk thresholds, assign priorities to the identified chunks, and re-replicate the identified chunks based substantially on the assigned priorities. The system may further monitor utilization of the servers, determine whether to redistribute any of the replicas, select one or more of the replicas to redistribute based on the utilization of the servers, select one or more of the servers to which to move the one or more replicas, and move the one or more replicas to the selected one or more servers.

Systems and methods for replicating data

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for modifying a computational graph to include send and receive nodes. Communication between unique devices performing operations of different subgraphs of the computational graph can be handled efficiently by inserting send and receive nodes into each subgraph. When executed, the operations that these send and receive nodes represent may enable pairs of unique devices to conduct communication with each other in a self-sufficient manner. This shifts the burden of coordinating communication away from the backend, which affords the system that processes this computational graph representation the opportunity to perform one or more other processes while devices are executing subgraphs.

Modifying computational graphs

A programming model provides a method for executing a program in a distributed architecture. One or more first shards of the distributed architecture execute one or more operations, and sending tuples to at least one second shard, the tuples being part of a stream and being based on the one or more operations. The one or more first shards send a token value to the at least one second shard when the sending of the tuples in the stream is complete. The at least one second shard determines whether a total of the token values matches a number of the one or more first shards, and takes a first action in response to determining that the total of the token values matches the number of the one or more first shards. The first action may include marking the stream as being complete and/or generating a message indicating that the stream is complete.

Method of executing a tuple graph program across a network

Growing community of inventors

Mountain View, CA, United States of America

Sanjay Ghemawat