Dragonfly Interconnect Adaptive Routing via Congestion Sensing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Processor interconnect networks in multiprocessor systems face inefficiencies due to high latency and cost associated with large numbers of global channels, especially in high-radix topologies, where conventional adaptive routing algorithms degrade performance by relying on local queue occupancy for congestion sensing, leading to suboptimal load balancing and increased latency.
Innovation Solution
The dragonfly topology employs high-radix routers with virtual channels and credit-round trip latency to sense global channel congestion, allowing for adaptive routing decisions based on remote information, reducing the number of global channels and increasing their length to minimize cost and latency, while using selective virtual-channel discrimination to balance load across channels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional adaptive routing algorithms use local queue occupancy for congestion sensing, then routing decisions can be made locally, but network latency increases and load balancing becomes suboptimal
Solution Approach 1:
The patent introduces an intermediary mechanism where routers exchange congestion information with their neighbors through dedicated congestion sensing links. This intermediary approach allows routers to make informed routing decisions based on remote congestion states without directly sensing every queue, thereby reducing latency while improving load balancing effectiveness.
Solution Approach 2:
The patent adds a new dimension to congestion sensing by introducing separate congestion sensing links that operate parallel to data transmission channels. This dimensional separation allows congestion information to be gathered and propagated independently from data traffic, enabling faster and more accurate routing decisions without contending for bandwidth.
2Productivity
If the number of global channels is increased to improve network capacity, then more processors can be connected, but network cost increases significantly
Solution Approach 1:
The patent segments the network into local and global channels, where local channels handle traffic within proximity groups and global channels handle long-distance communication. This segmentation reduces the total number of global channels needed while maintaining network capacity, as local traffic is handled efficiently by cheaper local connections.
Solution Approach 2:
The patent applies local quality by providing different channel characteristics for different communication needs: fast local channels for nearby processors and optimized global channels for distant communication. This differentiated approach reduces overall network cost by matching channel capacity to actual traffic requirements rather than uniformly provisioning all connections.
3Device complexity
If high-radix routers are used to reduce the number of router ports, then device complexity decreases, but latency increases due to larger switching fabric
Solution Approach 1:
The patent segments the routing function by using smaller radix routers organized in proximity groups rather than requiring a single large high-radix router. This segmentation reduces the switching fabric size at each router, lowering latency while achieving equivalent connectivity through coordinated group operation.
Solution Approach 2:
The patent adds a group dimension to the routing architecture, organizing routers into proximity groups that communicate locally before accessing global channels. This dimensional organization reduces the port radix requirement at individual routers while maintaining overall network connectivity and reducing switching latency.
Data Source
AI summary
A multiprocessor computer system comprises a dragonfly processor interconnect network that comprises a plurality of processor nodes and a plurality of routers. The routers are operable to adaptively route data by selecting from among a plurality of network paths from a target node to a destination node in the dragonfly network based on one or more of network congestion information from neighboring routers and failed network link information from neighboring routers.