Google Inc.

Cisco Technology, Inc.

Nandita Dukkipati

Gautam Kumar

Prashant R Chandra

Yuliang Li

Keon Jang

Inho Cho

Amin Vahdat

Hassan Wassel

Vytautas Valancius

Flavio Giovanni Bonomi

Weihuang Wang

Yi Cui

Vijaynarayanan Subramanian

Sangtae Ha

Hassan Mohamed Gamal Hassan Wassel

David Wetherall

Neelesh Bansod

Masoud Moshref Javadi

Michael Roger Marty

Srinivas Vaduvatha

Carlo Contavalli

Xiaoming Wang

Nathaniel Paul Lewis

Nicholas McDonald

Behnam Montazeri

Arjun Singhvi

Abhiram Ravi

Ahmed Said Mohamed Tawfik Issa

Praveen Kumar

Yaogong Wang

Xian Wu

Li Shi

Vinh The Lam

Serhat Arslan

Kirill Mendelev

Yiwen Zhang

Nicholas Mcdonald

Aspects of the disclosure provide for congestion control mechanisms to reduce data transmission queues and to increase link utilization through precise congestion signals and reduction of control loop delay. A congestion control system (CC) system can utilize Back-To-Sender (BTS) packets over (round trip time) RTT control loop delay to react to congestion faster. The CC system can apply Proactive Ramp-up (PRU) to identify flow completions to occupy released bandwidth right on time, e.g., as bandwidth is made available. The CC system can perform supply matching (SM) through network calculus concepts to increase link utilization. The CC system can apply some or all of the use of BTS packets, PRU, and/or SM to reduce network latency and improve data flow completion time as compared with other approaches.

Network congestion control in sub-round trip time

The technology is directed to the use of a bitmap generated at a receiver to track the status of received packets sent by a transmitter. The technology may include a network device including an input port, output port, and circuitry. The circuitry may generate a transmitter bitmap that tracks each data packet sent to another network device. The circuitry of the network device may receive, from the other network device, a receiver bitmap that identifies each data packet that is received and not received from the network device. The circuitry may then determine which data packets to retransmit by comparing the transmitter bitmap to the receiver bitmap.

Early retransmission for hardware reliable transport

Congestion control by adding a congestion signal tag header to each of one or more transmission packets prior to transmission of the transmission packets by the first node to a second node, the congestion signal tag header specifying one or more congestion signal types and, for each of the congestion signal types, specifying a congestion signal value by providing an initial congestion signal value for the congestion signal value; receiving one or more return packets generated by the second node in response to receipt of the transmission packets, the return packets including a congestion signal reflection header having one or more return congestion signal values, and the return congestion signal values corresponding respectively to the congestion signal types; and determining whether transmission rate control is necessary based on the return congestion signal values.

Network congestion control signaling on network devices

A distributed sender driven Admission Control System (ACS) is described herein, leveraging Weighted-Fair Quality of Service (QoS) queues, found in standard NICs and switches, to guarantee RPC level latency service level objectives (SLOs) by a judicious selection of QoS weights and traffic-mix across QoS queues. ACS installs cluster-wide RPC latency SLOs by mapping LS RPCs to higher weight QoS queues, and coping with overloads by adaptively apportioning LS RPCs amongst QoS queues based on measured completion times for each queue. When the network demand spikes unexpectedly to predetermined threshold percentage of provisioned capacity, ACS achieves a latency SLO that is significantly lower than the state-of-art congestion control at the 99.9th-p and admits significantly more RPCs meeting SLO target when RPC sizes are not aligned with priorities.

Admission control for latency-critical remote procedure calls in datacenters

A system includes a first processor configured to analyze packets received over a communication protocol system and determine one or more congestion indicators from the analysis of the data packets, the one or more congestion indicators being indicative of network congestion for data packets transmitted over a reliable transport protocol layer of the communication protocol system. The system also includes a rate update engine separate from the packet datapath and configured to operate a second processor to receive the determined one or more congestion indicators, determine one or more congestion control parameters for controlling transmission of data packets based on the received one or more congestion indicators, and output a congestion control result based on the determined one or more congestion control parameters.

Rate update engine for reliable transport protocol

A network interface card with traffic shaping capabilities and methods of network traffic shaping with a network interface card are provided. The network interface card and method can shape traffic originating from one or more applications executing on a host network device. The applications can execute in a virtual machine or containerized computing environment. The network interface card and method can perform or include several traffic shaping mechanisms including, for example and without limitation, a delayed completion mechanism, a time-indexed data structure, a packet builder, and a memory manager.

Fine grain traffic shaping offload for a network interface card

A system is provided for synchronizing clocks. The system includes a plurality of devices in a network, each device having a local clock. The system is configured to synchronize the local clocks according to a primary spanning tree, where the primary spanning tree has a plurality of nodes connected through a plurality of primary links, each node of the plurality of nodes representing a respective device of the plurality of devices. The system is also configured to compute a backup spanning tree before a failure is detected in the primary spanning tree, wherein the backup spanning tree includes one or more backup links that are different from the primary links. As such, upon detection of a failure in the primary spanning tree, the system reconfigures the plurality of devices such that clock synchronization is performed according to the backup spanning tree.

Fault tolerant design for clock-synchronization systems

Systems and methods for controlling congestion in a data network are provided. A base target round-trip time (RTT) for packets of a network flow including packets transmitted from a source network device to destination network device is obtained. A number of hops packets associated with the network flow traverse between the source network device and the destination network device is determined. A topology scaled target RTT for the network flow is determined based on the base target RTT and the determined number of hops. A congestion window size for the network flow is managed based on the topology scaled target RTT.

Congestion control for low latency datacenter networks

Systems and methods for controlling congestion of a data network are provided. An engine round-trip time (RTT) and a fabric RTT for a network flow are determined. An engine-based congestion window size for the flow is determined based on the engine RTT and a target engine RTT. A fabric-based congestion window size for the flow is determined based on the fabric RTT and a target fabric RTT. The smaller of the engine-based congestion window size and the fabric-based window size is selected for use in transmitting a future packet associated with the flow. The target engine RTT is determined based in part on the current congestion window used to transmit packets for the flow and/or the target fabric RTT is determined based on a number of hops packets associated with the flow traverse from a source to a destination associated with the flow.

A method for controlling congestion in a datacenter network or server is described. The server includes a processor configured to host a plurality of virtual machines and an ingress engine configured to maintain a plurality of per-virtual machine queues configured to store received packets. The processor is also configured to execute a CPU-fair fair queuing process to control the processing of the packets by the processor. The processor is also configured to selectively trigger temporary packet per second packet transmission limits on top of a substantially continuously enforced bit per second transmission limit upon detection of a per virtual machine queue overload.

Predictable virtualized NIC

A distributed bandwidth allocation system a distributed bandwidth limiter, a first throttler, and a second throttler. The distributed bandwidth limiter receives first usage data for a first entity and usage data for a second entity. Based on the first usage data, second usage data, and a total bandwidth allocation, the distributed bandwidth limiter determines a first bandwidth allocation specifying bandwidth available for network traffic for the first entity and a second bandwidth allocation that specifies bandwidth available for network traffic for the second entity, wherein a sum of the first bandwidth allocation and the second bandwidth allocation does not exceed the total bandwidth allocation. The first bandwidth allocation and the second bandwidth allocation are provided to respective throttlers than manage traffic for the first and second entities.

Distributed bandwidth allocation and throttling

Growing community of inventors

Menlo Park, CA, United States of America

Nandita Dukkipati