Akamai Technologies, Inc.

Google Inc.

Maxtor Corporation

Daniel Stodolsky

William Edward Weihl

Jay Parikh

Andrew Thomas Davis

Srinivasan Pichai

Andrew F Champagne

Michael M Afergan

Gregory Fletcher

Ashis Tarafdar

Shannon T Bailey

Ross Cohen

Nate Kushman

Mark Tsimelzon

Eddie Ruvinsky

Norman Paul Jouppi

Martin Guy Dixon

Parthasarathy Ranganathan

Quoc V Le

Sridhar Lakshmanamurthy

Sheng Li

Liqun Cheng

Satish L Rege

Tomer Margolin

R Hugo Paterson

Frederic R Carlson, Jr

Erik Karl Norden

Srinivasn Pichai

Aspects of the disclosure are directed to a heterogeneous machine learning accelerator system with compute and memory nodes connected by high speed chip-to-chip interconnects. While existing remote/disaggregated memory may require memory expansion via remote processing units, aspects of the disclosure add memory nodes into machine learning accelerator clusters via the chip-to-chip interconnects without needing assistance from remote processing units to achieve higher performance, simpler software stack, and/or lower cost. The memory nodes may support prefetch and intelligent compression to enable the use of low cost memory without performance degradation.

Heterogeneous ML accelerator cluster with flexible system resource balance

Implementations disclose leveraging aggregated network statistics for enhancing quality and user experience for live video streaming from mobile devices. A method includes receiving, by a processing device of a client device, a bandwidth parameter corresponding to aggregated network statistics for at least one of a current geographic location of the client device or a current network of the client device, initializing an upload quality parameter of an upload session based on the received bandwidth parameter, the upload session comprising upload of content from the client device, and modifying, by the processing device, the upload quality parameter throughout the upload session based on updated bandwidth parameters corresponding to aggregated network conditions for at least one of new geographic locations of the client device or new networks of the client device, the upload quality parameter to control a format of the upload session.

Leveraging aggregated network statistics for enhancing quality and user experience for live video streaming from mobile devices

The present invention leverages an existing content delivery network infrastructure to provide a system that enhances performance for any application that uses the Internet Protocol (IP) as its underlying transport mechanism. An overlay network comprises a set of edge nodes, intermediate nodes, and gateway nodes. This network provides optimized routing of IP packets. Internet application users can use the overlay to obtain improved performance during normal network conditions, to obtain or maintain good performance where normal default BGP routing would otherwise force the user over congested or poorly performing paths, or to enable the user to maintain communications to a target server application even during network outages.

Reliable, high-throughput, high-performance transport and routing mechanism for arbitrary data flows

An application deployment model for enterprise applications to enable applications to be deployed to and executed from a globally distributed computing platform, such as an Internet content delivery network (CDN). According to the invention, application developers separate their Web application into two layers: a highly distributed edge layer and a centralized origin layer. In a representative embodiment, the edge layer supports a servlet container that executes a Web tier, typically the presentation layer of a given Java-based application. Where necessary, the edge layer communicates with code running on an origin server to respond to a given request. In an alternative embodiment, the edge layer supports a more fully-provisioned application server that executes both Web tier (e.g., presentation) and Enterprise tier application (e.g., business logic) components. In either case, the inventive framework enables one or more different applications to be deployed to and executed from the edge server on behalf of one or more respective entities.

Java application framework for use in a content delivery network (CDN)

A method and system of load balancing application server resources operating in a distributed set of servers is described. In a representative embodiment, the set of servers comprise a region of a content delivery network. Each server in the set typically includes a server manager process, and an application server on which edge-enabled applications or application components are executed. As service requests are directed to servers in the region, the application servers manage the requests in a load-balanced manner, and without any requirement that a particular application server spawned on-demand.

Method of load balancing edge-enabled applications in a content delivery network (CDN)

The subject matter herein provides for a method and apparatus for comparison of network systems using live traffic in real-time. The inventive technique presents real-world workload in real-time with no external impact (i.e. no impact on the system under test), and it enables comparison against a production system for correctness verification. A preferred embodiment of the invention is a testing tool for the pseudo-live testing of CDN content staging servers, According to the invention, traffic between clients and the live production CDN servers is monitored by a simulator device, which then replicates this workload onto a system under test (SUT). The simulator detects divergences between the outputs from the SUT and live production servers, allowing detection of erroneous behavior. To the extent possible, the SUT is completely isolated from the outside world so that errors or crashes by this system do not affect either the CDN customers or the end users. Thus, the SUT does not interact with end users (i.e., their web browsers). Consequently, the simulator serves as a proxy for the clients. By basing its behavior off the packet stream sent between client and the live production system, the simulator can simulate most of the oddities of real-world client behavior, including malformed packets, timeouts, dropped traffic and reset connections, among others.

Method and apparatus for testing request-response service using live connection traffic

A method and system of load balancing application server resources operating in a distributed set of servers is described. In a representative embodiment, the set of servers comprise a region of a content delivery network. Each server is the set typically includes a server manager process, and an application server on which edge-enabled applications or application components are executed. As service requests are directed to servers in the region, the application servers manage the requests in a load-balanced manner, and without any requirement that a particular application server be spawned on-demand.

The present invention provides for a method and apparatus for comparison of network systems using live traffic in real-time. The inventive technique presents real-world workload in real-time with no external impact (i.e. no impact on the system under test), and it enables comparison against a production system for correctness verification. A preferred embodiment of the invention is a testing tool for the pseudo-live testing of CDN content staging servers, According to the invention, traffic between clients and the live production CDN servers is monitored by a simulator device, which then replicates this workload onto a system under test (SUT). The simulator detects divergences between the outputs from the SUT and live production servers, allowing detection of erroneous behavior. To the extent possible, the SUT is completely isolated from the outside world so that errors or crashes by this system do not affect either the CDN customers or the end users. Thus, the SUT does not interact with end users (i.e., their web browsers). Consequently, the simulator serves as a proxy for the clients. By basing its behavior off the packet stream sent between client and the live production system, the simulator can simulate most of the oddities of real-world client behavior, including malformed packets, timeouts, dropped traffic and reset connections, among others.

An application deployment model for enterprise applications to enable such applications to be deployed to and executed from a globally distributed computing platform, such as an Internet content delivery network (CDN). According to the invention, application developers separate their Web application into two layers: a highly distributed edge layer and a centralized origin layer. In a representative embodiment, the edge layer supports a servlet container that executes a Web tier, typically the presentation layer of a given Java-based application. Where necessary, the edge layer communicates with code running on an origin server to respond to a given request. In an alternative embodiment, the edge layer supports a more fully-provisioned application server that executes both Web tier (e.g., presentation) and Enterprise tier application (e.g., business logic) components. In either case, the inventive framework enables one or more different applications to be deployed to and executed from the edge server on behalf of one or more respective entities.

A data recording device including recording media having a set of at least two alternate regions thereon for each data segment, whereby each data segment has a corresponding set of at least two alternate regions on the recording media; a transducer positionable to write data on and read data from the recording media; and a controller for recording one or more data segments on the recording media in response to one or more write commands, the controller including a selector wherein for each of said one or more data segments the selector selects one region from the set of alternate regions corresponding to that data segment and positions the transducer to write that data segment on the selected region. To reduce the access time for each write operation, the selector selects the one region from the set of alternate regions based on shortest access time required to access each of the regions in the set of alternate regions for recording data.

Growing community of inventors

Somerville, MA, United States of America

Daniel Stodolsky