Amazon Technologies, Inc.

Wichorus, Inc.

Santosh Chandrachood

Linchi Zhang

Anthony A Virtuoso

Bijay Singh Bisht

Mehul A Shah

Pawan Uberoy

Aravind Ramarathinam

Amit Purohit

Ankit Kamboj

Vinayak Thapliyal

Ramkumar Kamalapuram Sugavanam

Min Zhou

Gayatri Ramesh Deo

Eric Mills

Lukasz Misiuda

Wesley Blumenthal

Nitin Bahadur

Rehan Jalil

Rahul Sharma Pathak

Pratik Bhagwat Gawande

Hei Tao Fung

Matt Bachman

Sampath Shreekantha

Chang Woo Lee

Harish Sitaraman

Dajung Kim

Child tasks of computational jobs are routed to cells of a network-based service. A request received at a request router to perform a child task of a computational job may be received. The request router may parse the request to extract a cell identifier for the run of the computational job. The cell identifier may then be used to identify the cell out of multiple cells implementing the network-based service. The cell identifier may be provided by the cell as part of performing a parent task of the run of the computational job. Then, the request may be sent to the identified cell to perform the child task of the run of the computational job.

Routing child tasks of computational jobs to cells of a network-based service

Techniques for self-tuning an analytics system via observed execution optimization are described. Upon a need for execution resources, a resource manager can select a type of executor from multiple candidate executor types based at least in part on one or more of current execution data associated with the execution of tasks of a user application and/or historic execution data associated with one or more other applications. The current execution data may include event log data originated by the driver application based on the execution of the user application and/or metric data describing characteristics of one or more worker nodes involved with executing the user application or characteristics of one or more other executors implemented by the one or more worker nodes in executing the user application.

Self-tuning analytics system with observed execution optimization

Techniques for a distributed data processing application service in a cloud provider network are described. A virtual machine bundle is obtained, the virtual machine bundle including a copy of a memory and one or more registers of a paused virtual machine executing a distributed data processing application runtime. A request to launch a virtual machine to execute a distributed data processing application runtime is received, the request to launch the virtual machine including a bundle identifier that identifies the virtual machine bundle. The virtual machine is launched based at least in part on the virtual machine bundle. The virtual machine sends a message to register the virtual machine with a cluster of virtual machines.

Distributed data processing application service in a provider network

Techniques for distributed data processing application service networking in a cloud provider network are described. A first request to launch a virtual machine is received, the first request including a session network identifier to identify a network of a distributed data processing application cluster hosted by the provider network. The first virtual machine is launched. An internet protocol (IP) address of the first virtual machine is set, and the IP address of the first virtual machine is formed at least in part by combining an IP subnet assigned to the computer system with the session network identifier. A firewall controlling network traffic to and from the first virtual machine is configured to allow packets having a source or destination address that matches a portion of the IP address of the first virtual machine that includes the session network identifier.

Distributed data processing application service networking in a provider network

In response to receiving a parallel processing job from a customer, a system operated by a computing resource service provider allocates and configures a cluster of computer systems capable of executing the job. In an embodiment, each computer system is configured with a first network stack that allows access to resources of the computing resource service provider and a second network stack that allows access to resources of the customer. In an embodiment, the state of the cluster is monitored by the system via the first network stack. In an embodiment, the system deploys a set of tasks on the cluster for fulfilling the processing job. In an embodiment, the tasks have access to the second network stack so that each task can access resources of the customer.

Cluster management

A system operated by a computing resource service provider maintains a pool of computing resources for use in processing job requests submitted by customers. The system allocates computing resources to the pool in anticipation of future resource needs. In various embodiments, future resource needs can be estimated based on scheduled jobs, or historical job information. In an embodiment, the computing resources are virtual computer systems which may be arranged in a cluster. In response to receiving a parallel processing job from a customer, the system reserves computing resources from the pool for performing the job. In an embodiment, the reserved resources are configured with a network namespace that is able to access to a customer's resources.

Resource management

Achieving resource management within a network processing unit (NPU) can be challenging due to the NPU's fixed and limited instruction set. In the case of large mobile networks, contexts are dynamic, and each subscriber in the network consumes multiple hardware resources, where a central processing unit (CPU) is typically employed to program individual resources even though generally impractical in high-session rates. To avoid programming individual resources, an example embodiment of the present invention creates a group of resources depending on a category, the group being shared dynamically among subscribers when a subscriber is active in the network. The example embodiment may include a network service processor (NSP) located in a forwarding path, such that the NPU may operate in connection with the NSP to allocate NPU resources using a packetized interface. The example embodiment achieves resource management within the NPU without a host CPU's involvement and without statically delaying resources.

Method and apparatus for assigning resources in a network node

In current hardware forwarding architectures, in which subscriber profiles are distributed across multiple cards in the network, lack of subscriber knowledge leads to loss of bandwidth on fabric links connecting multiple cards in a chassis, loss of device processing cycles, and loss of quality of service knowledge because the traffic is not classified. An example embodiment of the present invention employs a subscriber-aware switch programmed with a subscriber table that maps subscriber traffic to an 'anchor' line card with the subscriber's profile. The subscriber table allows for traffic to reach the anchor line card directly, avoiding unnecessary hops and loss of traffic information. As a result, line card, fabric, and other resources are used efficiently, thereby enabling a gateway or other network node to support high traffic rate (e.g., 100 Gbps) network models.

Method and apparatus for switching communications traffic in a communications network

In current systems, a typical way to collect application statistics includes sending requests to a resource manager that can access the resource hardware via a device driver. Current systems require multiple synchronous transactions between the processes, which results in the systems consuming large amounts of central processing unit resources that lead to sub-optimal rates of information retrieval. A method and apparatus configured to use asynchronous messaging across all modules and to return hardware statistics directly from the hardware to an application process, thereby bypassing transactions between the application and the resource manager, and bypassing similar transactions between the resource manager and a device driver. Embodiments of the invention are provided for minimizing the power consumed by the memory and minimizing the amount of dedicated memory necessary to perform.

Method and apparatus to report resource values in a mobile network

As part of mobility management messaging in current systems, only static quality of service (QoS) profiles and their statically configured classifiers are reported. These classifiers are known a priori and are typically used for hosted traffic. Once advanced application classifiers are matched and mobility messaging is done, the deep packet inspection (DPI) learned subscriber context is lost. An embodiment of the invention is configured to associate DPI-learned dynamic classifiers with the classifier's policy information as payload over mobility management messaging. The embodiment is useful for identifying protocols and applications, for authentication purposes, solving congestion-based issues between elements in a network, guaranteeing QoS without reserving resources statically, attaching traffic packets to a particular subscriber, and creating mobile signaling to provision end-to-end communication in a network. Thereby enabling the transfer of subscriber parameters and information to be associated with the subscriber throughout subscriber movement through multiple elements of a network.

Growing community of inventors

Milpitas, CA, United States of America

Santosh Chandrachood