Spine Node Data-Plane Bypass for Sub-Millisecond Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In spine-leaf architecture networks, link failures between spine and leaf nodes lead to packet loss and slow convergence due to reliance on the control plane for route convergence, resulting in flow interruptions lasting tens to hundreds of milliseconds.
Innovation Solution
A communication method and apparatus that enables fast bypass protection switching at the data plane by using a forwarding entry with an egress interface field and a fast bypass tag field to redirect traffic through a fast recovery leaf node group, achieving sub-millisecond recovery for CPU forwarding and microsecond-scale recovery for hardware chips.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the control plane performs route convergence after link failure, then routing correctness is maintained, but convergence speed is slow causing flow interruptions of tens to hundreds of milliseconds
Solution Approach 1:
The patent segments the forwarding decision process by separating the fast bypass tag verification (data plane) from the complete route convergence (control plane). The forwarding entry is divided into multiple components including the fast bypass tag field that can be independently checked without involving the control plane, enabling rapid failure detection and bypass while maintaining routing correctness through the slower control plane convergence.
Solution Approach 2:
The patent performs preliminary actions by pre-configuring fast bypass tags and alternative egress interfaces in the forwarding entries before link failures occur. When a link failure is detected, the spine node can immediately use the pre-prepared fast bypass tags to redirect traffic through alternative paths without waiting for control plane route convergence, thus achieving sub-millisecond recovery.
2Productivity
If traffic is forwarded through the failed downlink between spine1 and leaf3, then ECMP load balancing is maintained, but packet loss occurs
Solution Approach 1:
The patent implements feedback by having the spine node monitor the status of egress interfaces and detect link failures. When a failure is detected, the spine node uses the fast bypass tag to redirect traffic through alternative paths. This feedback mechanism ensures that traffic is automatically steered away from failed links while maintaining ECMP load balancing across healthy paths, preventing packet loss without manual intervention.
3Speed
If fast bypass protection switching is implemented at the data plane, then convergence speed increases to sub-millisecond, but device complexity increases due to additional forwarding entry fields and fast recovery leaf node group management
Solution Approach 1:
The patent merges the fast bypass protection switching functionality with the existing ECMP forwarding framework by integrating the fast bypass tag field into the forwarding entry structure. This allows the system to achieve sub-millisecond convergence speeds without requiring a completely separate fast failure recovery mechanism, thus limiting the increase in device complexity while still providing rapid failover capabilities.
Data Source
AI summary
In an example, a spine node may receive a first internet protocol (IP) packet sent by the first leaf through a first interface. The spine node may obtain a forwarding entry for forwarding the first IP packet; wherein, the forwarding entry comprises a fast bypass tag field and an egress interface field instructing a second interface for forwarding the first IP packet. The spine node may determine a third interface from a fast recovery leaf node group in response to determining that the second interface fails and the value of the fast bypass tag field is a first value, wherein the third interface is different from the second interface and is connected to a second leaf node. The spine node may send a second IP packet, which includes a destination media access control (MAC) address and the first IP packet, to the second leaf node through the third interface.