Disaggregated Switching Fabric Multicast Group Membership
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional data center networks face challenges with congestion, head-of-line blocking, inefficient routing, and scaling issues due to large and long-lasting traffic patterns, particularly in distributed computational workloads like machine learning and artificial intelligence, which require efficient bandwidth utilization and reduced latency.
Innovation Solution
The implementation of a disaggregated switching fabric with multicast hardware support allows for self-managed network group membership and configuration without software-based processing, utilizing spine-leaf configurations and hardware-based packet processing to distribute traffic efficiently and maintain network states dynamically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If software-based processing is used for network group membership management, then flexibility and control are improved, but processing latency and overhead increase
Solution Approach 1:
The patent extracts the network group membership management function from the software layer and implements it directly in the switching fabric hardware. This allows the membership management to occur at wire speed without software processing overhead, while still maintaining the necessary control and flexibility through programmable match-action pipelines in the hardware.
Solution Approach 2:
The patent replaces the mechanical/software-based packet processing system with a hardware-based switching fabric that uses match-action pipelines. This substitution enables packet processing to occur at line rate without the latency inherent in software-based processing, while maintaining programmability through the match-action framework.
2Productivity
If hardware-based packet processing is implemented, then processing speed and throughput are improved, but device complexity increases
Solution Approach 1:
The patent segments the packet processing function into discrete match-action pipelines within the switching fabric. Each pipeline independently processes packets based on specific match criteria and executes corresponding actions, allowing complex processing to be divided into manageable, parallel hardware components that can be configured dynamically.
Solution Approach 2:
The patent implements a universal match-action pipeline architecture that can handle multiple types of packet processing tasks through a single hardware framework. The same hardware infrastructure supports network group membership management, packet routing, quality of service policies, and other switching fabric functions, reducing overall system complexity despite the high processing capability.
3Productivity
If multicast hardware support is added to the switching fabric, then network group communication efficiency is improved, but manufacturing complexity increases
Solution Approach 1:
The patent merges the multicast functionality with the existing match-action pipeline architecture of the switching fabric. Instead of adding separate multicast hardware, the patent uses the programmable match-action pipelines to implement multicast group management and packet replication, leveraging the existing hardware resources to reduce manufacturing complexity.
Solution Approach 2:
The patent enables the switching fabric to self-configure multicast groups through hardware-based packet inspection and automatic group membership tracking. The match-action pipelines automatically learn and maintain multicast group information from observed traffic patterns, eliminating the need for complex external configuration systems and simplifying manufacturing.
Data Source
AI summary
Disclosed are systems, apparatuses, methods, and computer-readable media for configuring network groups without software-based processing and management. A method includes: transmitting a topology request from a first leaf node to a first spine node; in response to receiving a network topology reply from a second leaf node, generating a network topology identifying a group of leaf nodes, the network topology reply including network information associated with the second leaf node, and in response to a request to send a message from a first network node connected to the first leaf node to a second network node connected to the second leaf node, transmitting the message to the second leaf node based on the network information from the network topology reply. In some aspects, the first spine node or a second spine node are configured to receive the message and transmit the message to the second leaf node.