Multi-homed Fat-tree Routing in InfiniBand Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In InfiniBand fat-tree networks, existing routing algorithms fail to guarantee independent routes for multi-homed nodes, leading to potential single points of failure and performance bottlenecks due to reliance on physical port numbers and lack of redundancy in path selection.
Innovation Solution
The implementation of a multi-homed fat-tree routing algorithm (mFtree) ensures that each port on a multi-homed node takes mutually independent paths by using a subnet manager to mark and avoid shared switches, thereby providing redundancy and optimizing routing to prevent single points of failure and improve performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing routing algorithms are used in InfiniBand fat-tree networks, then routing is performed based on physical port numbers, but multi-homed nodes cannot guarantee independent routes leading to single points of failure
Solution Approach 1:
The routing algorithm segments the path selection process by treating each port on a multi-homed node as a separate routing entity. The mFtree algorithm divides the fat-tree network into multiple independent paths by tracking visited switches and preventing reuse across ports, ensuring route independence while maintaining manageable complexity through systematic path segmentation.
Solution Approach 2:
The routing algorithm performs preliminary actions by pre-marking switches as visited during the routing process for each port. This preliminary marking prevents subsequent ports from reusing the same switches, guaranteeing independent routes before potential failures can occur. The subnet manager proactively establishes these independent paths during network configuration and routing setup.
2Reliability
If physical port numbers are used for routing, then routing is simple to implement, but redundancy in path selection is lacking
Solution Approach 1:
The subnet manager acts as an intermediary between the physical port numbers and the routing decision-making process. It translates simple port identifiers into complex independent paths through the fat-tree network by tracking switch states and coordinating route selection across multiple ports, thereby providing path redundancy without requiring complex configuration at the endpoint devices.
Solution Approach 2:
The routing algorithm implements feedback mechanisms where the subnet manager monitors the state of switches and paths used by previous ports. This feedback information is used to make intelligent routing decisions for subsequent ports, selecting paths that have not been used by earlier ports to ensure independence and redundancy while maintaining ease of operation through automated decision-making.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and method can support multi-homed routing in a network environment, which can be based on InfiniBand architecture using a fat-tree or a similar topology. The system can provide an end node that is associated with a switch port on a leaf switch in a network fabric. Then, the system can perform routing for each of a plurality of ports on the end node, and ensure that the plurality of ports on the end node take mutually independent paths.