Modular Spine Switch Reduction for Simpler In-Network Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing in-network computing (INC) implementations face challenges due to the impracticality of adding additional hardware like memory and arithmetic logic units to leaf switches, leading to design complexity, congestion, and intricate error recovery in multi-tier topologies, especially when supporting data reduction collectives.
Innovation Solution
A modular switch architecture is introduced where data reduction operations are executed exclusively at the spine switch, utilizing line cards and a fabric element to handle data chunks and perform reduction operations, eliminating the need for hardware upgrades at leaf switches and simplifying error recovery.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If leaf switches are equipped with additional hardware (memory and arithmetic logic units) to support data reduction operations, then in-network computing capability is improved, but device complexity and cost increase
Solution Approach 1:
The patent extracts the data reduction functionality from leaf switches and relocates it to spine switches. Line cards at spine switches perform the reduction operations, while leaf switches maintain their simple forwarding function. This extraction resolves the contradiction by providing INC capability without adding complexity to leaf switches.
Solution Approach 2:
The patent shifts the reduction operation location from the network edge (leaf switches) to the network core (spine switches), changing the dimensional position of where reduction occurs. This allows leaf switches to remain simple while spine switches handle the complex reduction tasks.
2Adaptability or versatility
If multi-tier topology is used for INC reduction, then reduction capability is improved, but error recovery complexity and troubleshooting difficulty increase
Solution Approach 1:
The patent extracts the reduction operation from multiple tiers and consolidates it at the spine switch level. This creates a single-point reduction model that simplifies error recovery and troubleshooting compared to distributed multi-tier reduction.
3Loss of energy
If leaf switches perform reduction operations, then network bandwidth usage is reduced, but congestion is caused due to fixed packet paths
Solution Approach 1:
The patent implements dynamic path selection at spine switches for forwarding reduced results back to sources. Unlike fixed leaf switch paths, spine switches can adaptively route traffic based on current network conditions, reducing congestion while maintaining bandwidth efficiency.
Data Source
AI summary
Devices, systems, methods, and processes for in-network computing using modular switch architecture are described herein. Endpoint devices generate data chunks and forward them to a network, comprising spine and leaf switches, for data reduction. Leaf switches act as conduits and forward the data chunks to a spine switch. The spine switch includes various line cards (e.g., one for each leaf switch) and a fabric element. The line cards may execute a stage of data reduction on the received data chunks or may forward the received data chunks directly to the fabric element. The fabric element executes a data reduction operation on the data received from the line cards and obtains a reduced output which is forwarded to the endpoint devices via the line cards and the leaf switches. Thus, a single-tier in-network computing topology is implemented to execute data reduction in a cost-effective, simple, and efficient manner.