Clos Fabric Traffic Routing via Natural Traffic Splits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Weighted ECMP (WECMP) does not scale well in large networks due to resource limitations on network devices, as the weighting adds entries to forwarding tables that exceed memory capacity.
Innovation Solution
A system that leverages traffic splits naturally occurring in Clos fabrics to implement a WECMP-like result without additional entries in forwarding tables by abstracting sets of outgoing next hops across tier 1 devices, using ECMP or WECMP to manage capacity, and employing a monitoring server to dynamically adjust programming based on traffic patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Weighted ECMP (WECMP) is implemented to achieve precise traffic distribution and load balancing, then traffic engineering control is improved, but network device memory capacity is exceeded due to additional forwarding table entries
Solution Approach 1:
The patent segments the forwarding table into two distinct parts: a compressed ECMP representation that stores only essential routing information in a compact format, and a separate weight table that stores weighting information independently. This segmentation allows the system to maintain precise traffic distribution control through the weight table while keeping the compressed ECMP representation memory-efficient, thereby resolving the contradiction between traffic engineering control and memory capacity.
2Manufacturing precision
If traditional WECMP is used to achieve weighted routing outcomes, then traffic flow granularity is improved, but network device resource overhead increases
Solution Approach 1:
The patent extracts the weighting information from the traditional WECMP forwarding table entries and places it into a separate, dedicated weight table. This extraction allows the main compressed ECMP representation to maintain high traffic flow granularity without carrying the overhead of full WECMP entries. The system can still achieve precise weighted routing outcomes by combining the compressed representation with the separate weight table, thereby reducing network device resource overhead while preserving traffic flow granularity.
3Productivity
If WECMP entries are added to achieve desired load balancing, then load balancing performance is improved, but network scalability is reduced in large networks
Solution Approach 1:
The patent applies local quality by using a compressed representation for the ECMP routing information that is optimized for memory efficiency, while maintaining a separate weight table that provides the necessary weighting functionality only where needed. This localized approach allows the system to achieve good load balancing performance through the weight table without requiring full WECMP entries throughout the entire forwarding table, thereby improving network scalability in large networks while maintaining load balancing performance.
Data Source
AI summary
A computer-networking system is described to leverage traffic splits naturally occurring in Clos fabrics along the path to a top tier of the fabric. Using the traffic splits, a result similar to or the same as WECMP can be implemented but using less network-device overhead. The system introduces another level of indirection abstracting sets of outgoing next hops across tier 1 devices in the fabric and leveraging ECMP or WECMP to manage the capacity. Incoming traffic flows are split as they get forwarded across the fabric due to the fabric's topology, routing design, and/or the pattern of external connections. These traffic splits are leveraged in creating forwarding tables for the network devices, thereby making the scaling model for WECMP a function of the outgoing external connections of the fabric rather than the number of traffic flow destinations in the network domain.