Fat-Tree Routing via Hub Switch for Deadlock-Free IPoIB
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing routing algorithms in fat-tree topologies fail to provide deadlock-free and fully connected switch-to-switch communication, which is essential for efficient system management and diagnostics in high-performance computing clusters, particularly when using InfiniBand technology.
Innovation Solution
The implementation of a routing algorithm that selects a hub switch in a middleware machine environment to facilitate deadlock-free communication by creating a subtree with a root node, allowing all switches to communicate through this hub, and updating routing tables to ensure connectivity between switches that cannot reach each other directly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing routing algorithms are used in fat-tree topology, then network resource efficiency is improved, but deadlock-free and fully connected switch-to-switch communication is not achieved
Solution Approach 1:
The patent introduces a hub switch as an intermediary node to enable deadlock-free switch-to-switch communication. The hub switch acts as a mediator that receives packets from source switches and forwards them to destination switches, breaking the cyclic dependencies that cause deadlocks in traditional routing algorithms. This intermediary approach allows full connectivity while maintaining deadlock freedom.
Solution Approach 2:
The patent segments the fat-tree network into multiple subtrees, each rooted at a different switch. By dividing the network into these hierarchical segments, the routing algorithm can determine paths that avoid cyclic dependencies. The segmentation allows the system to maintain efficient resource utilization while preventing deadlocks through structured path selection.
2Device complexity
If traditional routing algorithms are used, then routing simplicity is maintained, but switch-to-switch communication connectivity is limited
Solution Approach 1:
The patent implements dynamic routing table updates at hub switches based on real-time network conditions and path availability. Instead of using static routing tables, the system continuously adapts routing information to ensure full connectivity between switches. This dynamic approach enables versatile switch-to-switch communication while maintaining manageable complexity through automated updates.
Solution Approach 2:
The hub switch is designed with multi-functionality, serving both as a regular network node and as a dedicated intermediary for deadlock-free routing. It performs standard switching functions while simultaneously managing routing decisions for multiple source-destination pairs. This universal design enhances connectivity without requiring separate dedicated routing infrastructure.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and method can support routing packets between a plurality of switches in a middleware machine environment, thereby supporting Internet Protocol (IP) based management traffic via enabling IP over Infiniband (IPoIB) communication in the middleware machine environment. The plurality of switches can perform routing for inter-switch traffic in the middleware machine environment using a routing algorithm. Then, a switch in the middleware machine environment can be selected as a hub switch for inter-switch traffic that can not reach destination using the routing algorithm. Furthermore, a routing table associated with the hub switch can be updated when a path exists between a source switch and a destination switch via the hub switch.