Dynamic Multipath Routing for High-Performance Computing Clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In high-performance computing clusters, static-routing networks like InfiniBand lead to link competition and decreased communication performance due to uneven use of communication paths, especially in fat tree configurations where packet transfer paths are not dynamically changed.
Innovation Solution
A multipath network with relay devices extending over two or more hierarchies is implemented, where each computer is assigned multiple network identifiers corresponding to different communication paths, allowing dynamic routing to distribute communication paths evenly across the network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If static-routing networks like InfiniBand are used, then network performance is improved through high-speed communication, but link competition occurs and communication performance decreases due to uneven use of communication paths
Solution Approach 1:
The patent implements dynamic routing by assigning multiple network identifiers (LIDs) to each computer, corresponding to different communication paths. The routing decision is made dynamically based on the destination computer's LID assignment, allowing packets to be routed to different paths depending on the communication pattern, thereby preventing link competition while maintaining high-speed communication
Solution Approach 2:
The patent segments the communication paths by assigning multiple LIDs to each computer, where each LID corresponds to a specific path through the network hierarchy. This segmentation allows independent routing decisions for different communications, distributing the load across multiple paths and preventing congestion on any single link
2Adaptability or versatility
If fat tree network configuration is used, then network scalability is improved, but link competition increases during all-to-all communication scenarios
Solution Approach 1:
The patent makes the routing dynamic by assigning multiple LIDs to each computer based on their position in the fat tree hierarchy. During all-to-all communication, the system dynamically selects appropriate LIDs to route packets through different paths, preventing concentration of traffic on specific links and maintaining high throughput despite network scalability
Solution Approach 2:
The patent applies local quality by assigning LIDs specifically based on each computer's hierarchical level and position in the fat tree structure. Computers at different levels receive tailored LID assignments that optimize their routing paths, ensuring that local communication patterns are optimized while maintaining overall network scalability
Data Source
AI summary
A computer acquires correspondence data including a plurality of combinations of identifiers of selected computers, in which parallel processes are started, and identifiers of the parallel processes, wherein the selected computers are among a plurality of computers connected to a multipath network having relay devices extending over two or more hierarchies. The computer specifies a network identifier to be used for distributing communication paths among the selected computers for every selected computer that includes an identifier in the correspondence data, wherein the specified network identifiers is among network identifiers of respective communication paths in the multipath network, and the specified network identifier corresponds to each identifier of a plurality of computers, and is used for routing.