Google Inc.

Jiangbo Li

Victor Lin

Junlan Zhou

Fei Ye

Zhengrong Ji

Subhasree Mandal

Joon Suan Ong

Ashish Naik

Ke Xiao

Junda Liu

Shidong Zhang

Qingxi Li

Systems and methods of hybrid control of a computer network are provided. The system can comprise a controller and a plurality of nodes having an in-band control plane network arranged in a directed acyclic graph routing topology. A first node of the plurality of nodes can be configured to detect a link failure of a first link, the first link being controller-directed. The first node can reverse a second link to a second node from outward-directed to controller-directed such that the first node can communicate with the controller via the second link and the in-band control plane network. The first node can detect a reversal of the second link caused by the second node. The first node can determine that the in-band control plane network is partitioned. The first node can use an out-of-band control plane network to inform the controller that the in-band control plane network is partitioned.

Systems and methods for hybrid control of computer networks

A system and method is provided for sending congestion notification messages through L3 networks. For example, a data packet is received at a first switch in a first fabric block of an L3 network, and the first switch performs source MAC tagging of the data packet. The data packet is then forwarded to a second switch in a second fabric block of the L3 network, and the source MAC tag is maintained by the second switch and any intermediate switches. The second switch determines, in response to receiving the data packet, whether it is congested, and generates a notification message if it is congested. The notification message is L2 forwarded to the first fabric block, and further forwarded from the first switch to a source of the data packet using ACL matching.

System and method for providing congestion notification in layer 3 networks

Methods and systems for generating a forwarding table for a packet switch. The system includes a route manager for the packet switch, configured to identify a plurality of multi-path groups each corresponding to a respective initial set of routing entries in the forwarding table and generate, for one or more multi-path groups, at least one replacement set of routing entries with fewer routing entries than the initial set corresponding to the respective multi-path group. The route manager selects, based on a traffic reduction cost metric, one or more of the replacement sets of routing entries, each corresponding to a different respective multi-path group, and updates the forwarding table with the selected replacement sets. In some implementations, the traffic reduction cost metric includes a traffic characteristic. In some implementations, the packet switch participates in a software-defined network (SDN) and the route manager is part of an SDN controller.

Prefix-aware weighted cost multi-path group reduction

Aspects of the present disclosure relate to detecting and repairing permanently pauses on a flow controlled fabric. In one aspect, one or more computing devices, such as a switch or a centralized controller, may detect whether a port of a network device receives one or more pause messages. The pause messages may instruct the network device to pause data transmission. Further, the one or more computing devices may determine a period of time during which the port receives the one or more pause messages and identify the port as a permanently paused port based on the determined period of time. The one or more computing devices may then reconfigure the permanently paused port to stop complying with the one or more pause messages.

Detection and repair of permanent pause on flow controlled fabric

This disclosure provides systems, methods, and apparatus for providing a network switch for implementing weighted cost multi-pathing (WCMP). In some implementations, the switch can implement equal cost multi-pathing (ECMP) using ECMP forwarding tables. The ECMP forwarding table may include more than one output port associated with a destination address of an incoming packet. The ECMP forwarding table may become undesirably large if unequal weights were assigned to the output ports. A multi-level ECMP table can be implemented that can allow assigning weights to the output ports while being relatively small. The multi-level ECMP table can include a first level where each entry includes at most a single output port. A second level can include entries having one or more output ports.

Systems and methods for implementing weighted cost multi-path using two-level equal cost multi-path tables

Systems and methods for managing a large number of multi-stage switches in a data center are provided. In one aspect, a plurality of first stage switches in a switch cluster may be configured to function as a proxy switch for a given number of second stage switches in the switch cluster. Each of the first stage switches may thus receive, via a first set of networking links, inbound control plane management traffic destined for respective ones of the given number of second stage switches. The first stage switches may then route the control plane management traffic to the appropriate second stage switches via a second set of bidirectional links. In addition, each of the second stage switches may be further configured to route outbound data traffic originating from the second stage switches using all available uplinks in the second set of bidirectional links.

Inband management in a multi-stage CLOS network

A method of testing network device connectivity in a network by a host computing device includes identifying a topology of the network, where the network includes a plurality of communication paths connecting a plurality of network devices. The method further includes developing a test procedure based on the topology and a plurality of constraints, where a first constraint in the plurality of constraints includes sending a test data stream through each communication path in the plurality of communications paths only once in each direction. The method further includes configuring the plurality of network devices to perform the test procedure, initiating the test procedure, where initiating includes sending the test data stream from the host computing device to a first network device in the plurality of network devices, and monitoring the progress of the test procedure.

Growing community of inventors

Sunnyvale, CA, United States of America

Jiangbo Li