Link Failure Handling for AI Cluster Traffic Forwarding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional ECMP traffic balancing mechanisms lead to uneven traffic distribution on links, causing some links to be congested or experience packet loss while others remain idle, resulting in reduced network throughput and lower service performance in AI clusters.
Innovation Solution
A network-wide load balancing solution based on computing-network collaboration, utilizing a dual-layer forwarding protection mechanism that includes intelligent routing and dynamic routing protocols to handle link failures, ensuring uninterrupted traffic flow and optimizing paths through monitoring and sensing mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If ECMP traffic balancing mechanism is used, then traffic distribution is simplified, but traffic distribution becomes uneven causing link congestion and packet loss
Solution Approach 1:
The patent segments the traffic balancing approach by introducing per-flow path selection instead of uniform ECMP hashing. Each flow is assigned to a specific path based on destination address, separating the traffic management function into flow-level decisions rather than link-level random distribution, thereby avoiding congestion on specific links while maintaining simplicity in implementation
Solution Approach 2:
The patent implements dynamic path selection where the forwarding decision changes based on the destination network address. The system dynamically determines the next hop and egress interface for each flow, allowing flexible adaptation to network topology changes and avoiding static ECMP hashing limitations, thus improving both simplicity and reliability
2Device complexity
If traditional ECMP hashing is used, then routing decision is simple, but some links become idle while others experience full throughput and congestion
Solution Approach 1:
The patent applies local quality by making forwarding decisions tailored to each flow's destination address rather than applying uniform ECMP hashing to all traffic. Each flow receives customized path selection based on its specific destination, optimizing local traffic distribution while maintaining simple routing logic at each network apparatus
Solution Approach 2:
The patent changes the routing parameter from random ECMP hash values to destination address-based path selection. By using the destination network address as the key parameter for forwarding decisions, the system achieves more efficient traffic distribution across links, improving overall network throughput while keeping the routing mechanism simple and implementable
3Ease of manufacture
If link failure handling uses traditional mechanisms, then implementation is straightforward, but packet loss persists for long periods
Solution Approach 1:
The patent implements preliminary action by having network apparatuses proactively monitor link status and pre-configured to switch to backup paths upon detecting failures. The system prepares alternative routing paths in advance and automatically activates them when link failures occur, eliminating lengthy packet loss periods while maintaining straightforward implementation through standard monitoring and switching mechanisms
Solution Approach 2:
The patent employs feedback mechanisms where network apparatuses continuously monitor link status and respond to failures by deleting failed flow entries and switching to dynamic routing protocols. This feedback loop ensures rapid detection and response to link failures, minimizing packet loss duration while keeping the implementation practical through standard networking protocols and monitoring tools
Data Source
AI summary
The present disclosure relates to a technical field of intelligent computing centers, and particularly to methods and apparatuses for handling a link failure, and storage mediums. The method includes: receiving a first flow entry sent by a controller, where the first flow entry is for forwarding the target flow, and the first flow entry includes a source network address, a destination network address, an egress interface, and information of a next hop apparatus of the target flow; monitoring a status of a link associated with the first flow entry; deleting the first flow entry in response to detecting that the link associated with the first flow entry faults; and forwarding the target flow based on the destination network address of the target flow and a dynamic routing entry.