Multi-Stage Switch Fabric for Low Latency Data Center Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data center systems face high costs and latency due to complex architectures with multiple layers of switches, leading to inefficient scalability and increased end-to-end latency, as well as poor handling of congestion and store-and-forward delays.
Innovation Solution
A multi-stage switch fabric with a switch core that provides any-to-any connectivity and low latency, eliminating the need for repeated packet processing and queuing at each layer, and implementing a request-grant scheme for cell-based flow control to prevent loss and minimize latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple layers of switches are used for packet processing, then connectivity and routing capabilities are improved, but latency and system cost increase
Solution Approach 1:
The patent merges multiple layers of packet processing into a single unified switch fabric architecture. Instead of having separate switching layers that repeat processing functions, the invention combines these functions into one cohesive system that maintains connectivity capabilities while eliminating redundant processing steps, thereby reducing latency.
Solution Approach 2:
The switch fabric is designed to perform multiple functions simultaneously - routing, switching, and packet processing - within a single unified architecture. This multi-functional approach eliminates the need for separate processing layers while maintaining the adaptability and connectivity capabilities previously achieved through layered structures.
2Adaptability or versatility
If multiple layers of switches are added for scalability, then network capacity is improved, but cost and complexity increase
Solution Approach 1:
The switch fabric is divided into modular components including input ports, output ports, and intermediate stages that can be independently configured and scaled. This segmentation allows the system to achieve high network capacity through modular expansion rather than adding complex layered structures, thereby reducing overall system complexity and cost.
Solution Approach 2:
The patent transitions from a vertical layered architecture to a horizontal planar switch fabric layout. This dimensional change allows for more efficient space utilization and simpler signal routing, enabling scalability without proportionally increasing complexity or cost.
3Reliability
If store-and-forward processing is used at each layer, then packet processing reliability is improved, but zero-load latency increases
Solution Approach 1:
The switch fabric performs packet processing actions in advance as packets enter the fabric, rather than waiting for complete packet reception at each layer. This preliminary processing approach maintains reliability by ensuring packets are properly routed and formatted early in the transmission process, while eliminating the store-and-forward delay that would otherwise occur at multiple layers.
4Adaptability or versatility
If conventional packet processing is repeated at each layer, then routing flexibility is improved, but processing overhead and cost increase
Solution Approach 1:
The unified switch fabric implements intelligent routing decisions based on feedback from the packet header and destination information. Instead of repeating processing at each layer, the system makes routing decisions once at the fabric level using feedback about packet characteristics, thereby maintaining routing flexibility while significantly reducing processing overhead.
Data Source
AI summary
In one embodiment, an apparatus includes a switch core that has a multi-stage switch fabric. The multi-stage switch fabric has a set of ingress ports and a set of egress ports. The switch core can be configured to be coupled to a set of edge devices via the set of ingress ports and the set of egress ports. The switch core can be configured to receive a packet from an ingress port from the set of ingress ports. The switch core can be configured to send a set of cells associated with the packet from the ingress port to an egress port from the set of egress ports without a store-and-forward delay associated with a zero-load latency for the switch core.