EVPN Fabric Congestion Control With Ingress VO Queues
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional network fabrics experience significant congestion due to the lack of explicit awareness of egress port availability and the need for determining appropriate egress ports and tagging operations at the egress switch, leading to processing delays and reduced efficiency and performance of applications running on GPUs.
Innovation Solution
Implementing a disaggregated scheduled fabric (DSF) with an Ethernet Virtual Private Network (EVPN) control plane, utilizing egress managers, ingress managers, and virtual output (VO) queues to manage congestion by storing data packets at the ingress switch until the egress switch is available, and including necessary identifiers and encapsulation values in the data packet headers for smooth processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the ingress switch forwards data packets without explicit awareness of egress port availability, then the forwarding process is simple and fast, but significant congestion occurs in the network fabric
Solution Approach 1:
The ingress switch performs preliminary actions by determining the egress port and creating the VO queue before forwarding data packets. This advance preparation eliminates the need for egress switch lookup operations, reducing processing delays and preventing congestion in the network fabric while maintaining simple forwarding operations
2Reliability
If the egress switch determines the appropriate egress port and tagging operations, then the data packets are correctly routed, but processing delays are introduced
Solution Approach 1:
The patent extracts the egress port determination function from the egress switch and relocates it to the ingress switch. The ingress switch now performs the lookup and creates VO queues in advance, eliminating the time-consuming egress switch processing steps while ensuring accurate routing through pre-computed forwarding information
3Productivity
If numerous data packets are stuck in the network fabric at spine switches and egress switches, then congestion occurs, but the traditional forwarding approach without VO queues cannot prevent this
Solution Approach 1:
The ingress switch performs preliminary actions by determining the egress port and creating the VO queue before forwarding data packets. This advance preparation eliminates the need for egress switch lookup operations, reducing processing delays and preventing congestion in the network fabric
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Devices, systems, methods, and processes for fabric congestion management are described herein. At each ingress switch, virtual output ("VO") queues are created for egress ports based on identifiers, state indicators, and encapsulation values of the egress ports received via an Ethernet Virtual Private Network ("EVPN") control plane. When a data packet is received at the ingress switch, an egress port for the data packet is determined, an identifier and an encapsulation value of the egress port are added to the data packet, and the data packet is stored in a corresponding VO queue. The data packet remains at the ingress switch until an egress switch is available. At the egress switch, one or more tags are added in the data packet based on the encapsulation value, whereas the destination egress port is identified based on the identifier. Thus, a quick egress through the egress switch is achieved.