Weighted Fat-Tree Routing for HPC Network Load Balancing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy fat-tree routing algorithms in high-performance computing (HPC) environments fail to effectively balance network load due to ignoring traffic characteristics of nodes, leading to sub-optimal throughput and unpredictable performance, as they assume equal weights for all nodes and rely on indexing order for route assignment.
Innovation Solution
The weighted fat-tree routing algorithm (wFatTree) assigns receive weights to end nodes based on their traffic profiles, sorting them by decreasing receive weights and updating weights on ports to prioritize high-traffic nodes, ensuring balanced load distribution across the network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If legacy fat-tree routing algorithms assume equal weights for all nodes and rely on indexing order for route assignment, then the routing process is simple and deterministic, but network throughput is sub-optimal and performance is unpredictable due to ignoring traffic characteristics
Solution Approach 1:
The patent introduces receive weights as a new parameter to characterize traffic characteristics of end nodes. By changing from uniform weight assumption to variable weight parameters based on actual traffic profiles, the routing algorithm achieves optimal throughput while maintaining manageable complexity through systematic weight assignment and accumulation mechanisms
Solution Approach 2:
The patent performs preliminary sorting of end nodes by receive weight before route assignment. This preliminary action organizes the routing process to prioritize high-traffic nodes first, ensuring that routes are assigned in an order that maximizes network throughput from the beginning rather than requiring complex real-time adjustments
2Productivity
If legacy fat-tree routing algorithms use indexing order for route assignment, then the routing process is deterministic and easy to implement, but load balancing is ineffective and network congestion occurs
Solution Approach 1:
The patent replaces the simple indexing order parameter with receive weight parameters that reflect actual traffic characteristics. This parameter change enables effective load balancing by routing decisions based on node importance rather than arbitrary index values, while the systematic weight accumulation process keeps implementation complexity manageable
Solution Approach 2:
The patent incorporates feedback mechanisms where receive weights are determined based on actual traffic profiles of end nodes. This feedback loop allows the routing algorithm to adapt to real network conditions, improving load balancing effectiveness while the structured feedback process maintains operational simplicity
3Reliability
If weighted fat-tree routing algorithm prioritizes high-traffic nodes by sorting and weight accumulation, then network throughput improves and congestion reduces, but routing computation becomes more complex
Solution Approach 1:
The patent performs preliminary sorting of end nodes by receive weight and pre-calculates weight accumulations for different port selections. This preliminary computation organizes the data in advance, making the actual routing decisions more straightforward and improving performance predictability while containing computation complexity through structured preprocessing
Solution Approach 2:
The patent transforms the routing problem from simple index-based selection to weight-based selection with accumulation tracking. By changing the decision parameters to include receive weights and accumulated weights, the system achieves reliable and predictable performance while the systematic parameter management keeps computation complexity controlled
Data Source
AI summary
Methods and systems for supporting efficient load balancing among a plurality of switches and a plurality of end nodes arranged in a tree topology in a network environment. The methods and systems can sort the plurality of end nodes, wherein the plurality of end nodes are sorted in a decreasing order of a receive weight. The method and system may further route, in the decreasing order of receive weights, the plurality of end nodes, wherein the routing comprises selecting at least one down-going port and at least one up-going port. Further, the method and system can increase an accumulated downward weight on each selected down-going port by the receive weight of the routed end node, and increase an accumulated upward weight on each selected up-going port by the receive weight of the routed end node.