No relevant resource is found in the selected language.

This site uses cookies. By continuing to browse the site you are agreeing to our use of cookies. Read our privacy policy>Search

Reminder

To have a better experience, please upgrade your IE browser.

upgrade

Configuration Guide - Low Latency Network

CloudEngine 8800, 7800, 6800, and 5800 V200R005C00

This document describes the configurations of Buffer optimization of lossless queues, Fast ECN, Fast CNP, Dynamic ECN threshold of lossless queues, and Dynamic load balancing.
Rate and give feedback:
Huawei uses machine translation combined with human proofreading to translate this document to different languages in order to help you better understand the content of this document. Note: Even the most advanced machine translation cannot match the quality of professional translators. Huawei shall not bear any responsibility for translation accuracy and it is recommended that you refer to the English document (a link for which has been provided).
Overview of Low-latency Networks

Overview of Low-latency Networks

Definition

A low-latency network uses the following technologies to eliminate packet loss and achieve low latency on an Ethernet network:

  • Optimize the queue buffer threshold to ensure that packets of lossless queues are transmitted with no loss as much as possible.

    Assign the buffer threshold properly based on the chip's forwarding capability, ensuring lossless forwarding of packets of lossless queues.

  • Use fast Explicit Congestion Notification (ECN) to shorten the congestion notification time.

    When a queue is congested, the ECN flag is added for outgoing packets in the queue. This eliminates the time taken to add the ECN flag to a packet entering the queue before forwarding the packet with the ECN flag, which shortens the congestion notification time.

  • Use fast Congestion Notification Packet (CNP) to immediately adjust the packet sending rate of the source end.

    When congestion occurs on the forwarding device, the forwarding device replaces the destination server to send a CNP to the source server. The source server then adjusts the packet sending rate, relieving congestion of the queue buffer on the forwarding device.

  • Identify elephant and mice flows to ensure the forwarding latency of mice flows in lossless queues.

    Identify elephant and mice flows in lossless queues and preferentially schedule packets of mice flows to reduce the forwarding latency of mice flows.

  • Dynamically adjust the ECN threshold to balance the tradeoff between low latency and high throughput of lossless queues.

    Dynamically adjust the ECN threshold of lossless queues based on the incast concurrency and proportions of elephant and mice flows to balance the tradeoff between low latency and high throughput of lossless queues.

  • Use dynamic load balancing and select a least congestion link to forward packets.

    In the multi-path scenario, measure the congestion status of each link and select a least congestion link to forward packets.

Purpose

As data centers develop continuously, popular applications such as distributed storage, high-performance computing, and AI place increasingly stringent requirements on networks. Physical networks are required to provide lossless transmission, low latency, and high throughput, as well as being able to deploy services on a large scale.

Traditionally, the InfiniBand (IB) network is used to achieve mutual access between high-performance computing resources, and the Fibre Channel (FC) network is used to implement interworking between computing and storage resources. The cost of IB and FC networks is two to three times higher than that of Ethernet networks. As 25G, 50G, and 100G Ethernet networks are becoming increasingly mature, Ethernet networks are starting to replace IB and FC networks due to their low cost.

Traditional Ethernet networks use best-effort forwarding, which means they cannot ensure low latency or low packet loss. Most traffic of popular applications such as distributed storage, high-performance computing, and AI is processed according to the many-to-one incast model. When the traffic is forwarded on Ethernet networks, congestion and packet loss easily occur on network devices. In addition, elephant and mice flows are often transmitted together. The mice flows cannot be scheduled immediately due to elephant flows, resulting in a long delay.

The low-latency network resolves problems of burst packet loss and long delay on traditional Ethernet networks, enabling Ethernet networks to meet the high performance requirements of distributed storage, high-performance computing, and artificial intelligence applications.

Basic Concepts

Lossless service and lossy service

Services are classified as either lossless or lossy services, depending on whether packets need to be transmitted with no loss:
  • Lossless service: indicates a service of which packets need to be transmitted with no loss.

    Lossless transmission of packets on the Ethernet network is implemented through Priority-based Flow Control (PFC). The device sets the PFC threshold for eight queues on each port. When the buffer used by a queue exceeds the PFC threshold, the device sends a PFC notification packet to the upstream device, requesting the upstream device to stop sending packets. When the buffer used by a queue falls below the PFC threshold, the device sends a PFC stop packet to the upstream device, requesting the upstream device to send packets again. By doing this, packets are transmitted with no loss.

    The device supports PFC based on 802.1p and DSCP priorities.
    • PFC based on 802.1p priorities: An 802.1p priority corresponds to one interface queue on the device. For example, 802.1p priority 0 corresponds to interface queue 0 and 802.1p priority 1 corresponds to interface queue 1, and so on.
    • PFC based on DSCP priorities: The device maps DSCP priorities to internal priorities based on the configured DiffServ domain, and an internal priority corresponds to one interface queue. For details, see "Priority Mapping Configuration" in CloudEngine 8800, 7800, 6800, and 5800 Seriesswitch Configuration Guide - QoS.

    When the device is enabled with PFC, the configured 802.1p priority or internal priority is the lossless priority and the queue corresponding to the priority is the lossless queue.

  • Lossy service: indicates the service that permits packet loss. When the device is not enabled with PFC, the configured 802.1p priority or internal priority is the lossy priority and the queue corresponding to the priority is the lossy queue.

Translation
Download
Updated: 2019-04-20

Document ID: EDOC1100040243

Views: 5541

Downloads: 100

Average rating:
This Document Applies to these Products
Related Documents
Related Version
Share
Previous Next