Multi-stage Switching Topology for Data Center Fabric Scalability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data center networking fabrics built using Clos architecture are costly and suffer from scalability and performance issues such as bandwidth stranding and oversubscription, while alternative architectures like flattened butterfly and Dragonfly face complexity and manageability challenges.
Innovation Solution
A multi-stage switching topology comprising multiple tiers of switches with interconnections between switching chips, forming building blocks and superclusters, allowing for efficient routing and load balancing through intra-switch, inter-switch, and inter-block connections, and utilizing a processor for traffic distribution and congestion management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Clos architecture is used to build data center networking fabrics, then fully connected non-blocking fabric is achieved, but cost and scalability are worsened due to high chip count and bandwidth stranding
Solution Approach 1:
The patent segments the Clos architecture into multiple hierarchical tiers (first tier switches, second tier switches, core switches) with controlled interconnections. Each tier handles specific traffic patterns, dividing the monolithic fabric into manageable segments that reduce overall chip count while maintaining non-blocking connectivity through structured path selection.
Solution Approach 2:
The patent introduces a hierarchical dimension to the network fabric by organizing switches across multiple tiers with systematic uplink/downlink patterns. This vertical stratification transforms the traditional flat Clos architecture into a multi-layered structure that reduces bandwidth stranding by directing traffic through optimal hierarchical paths.
2Device complexity
If alternative architectures like flattened butterfly and Dragonfly are used, then cost is reduced, but performance and manageability are worsened due to oversubscription and complexity
Solution Approach 1:
The patent implements dynamic routing capabilities within the hierarchical structure, allowing traffic to be routed flexibly across multiple tiers based on destination location and current network conditions. This dynamic path selection maintains performance by avoiding oversubscription while keeping the physical topology simpler and more manageable than alternative architectures.
Solution Approach 2:
The patent introduces intermediate tier switches as mediators between edge switches and core switches. These intermediate devices simplify manageability by providing localized routing intelligence and traffic aggregation, reducing the complexity burden on core switches while maintaining performance through distributed intelligence.
3Productivity
If multi-stage switching topology with multiple tiers is implemented, then scalability and routing efficiency are improved, but structural complexity increases
Solution Approach 1:
The patent applies a nested hierarchical structure where first tier switches are contained within building blocks, which are in turn connected to core switches forming larger data center fabrics. This nesting allows incremental scalability - individual building blocks can be deployed independently or combined - while routing efficiency is maintained through consistent hierarchical path selection at each nesting level.
Data Source
AI summary
A multi-stage switching fabric provides a first tier and a second tier. The first tier may include a first plurality of switches, and the second tier may include a second plurality of switches, wherein each of the switches in the first tier are coupled to the switches in the second tier. Each of the switches in the first and second tiers may include a plurality of switching chips, wherein a given switching chip in a given switch is coupled to each other switching chip in the given switch. The fabric may further include a third tier comprising one or more core switches, wherein each switch in the second tier is coupled to at least one of the core switches.