Quad Full Mesh Dimension-Driven Network Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data center network architectures face challenges in reducing the number of hops for data communication between nodes, leading to high latency and energy consumption, and struggle to efficiently scale and isolate faults.
Innovation Solution
A quad node full mesh and dimension-driven hyper-torus clustering architecture with a lightweight topology protocol is implemented, utilizing 3 links per node to create a full mesh connection across multiple dimensions, optimizing the number of hops and allowing for efficient scaling and fault isolation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If traditional network architectures are used to connect servers, then the network can support basic data communication, but the number of hops for data travel increases leading to high latency and energy consumption
Solution Approach 1:
The patent applies dimensionality change by organizing nodes into a multi-dimensional torus network architecture where nodes are positioned in N dimensions with coordinates (x1, x2, ..., xN). Each dimension provides direct communication paths, reducing the number of hops required for data travel compared to traditional flat network architectures. The dimensional structure allows packets to traverse multiple dimensions simultaneously, optimizing the path between source and destination nodes.
Solution Approach 2:
The network is segmented into discrete nodes and dimensions, with each node having specific coordinates in the N-dimensional space. The communication paths are segmented into hops along each dimension, allowing independent optimization of routing in each dimension. This segmentation enables the network to break down complex routing problems into simpler dimensional components.
2Speed
If the number of communication links per node is increased to reduce hops, then data transfer speed improves, but the device complexity and cost increase
Solution Approach 1:
The patent changes the parameter of network topology from traditional flat or hierarchical structures to an N-dimensional torus configuration. This parameter change allows each node to maintain a consistent, limited number of communication links (typically 2N links for N dimensions) while achieving reduced hop counts through the dimensional arrangement. The dimensional parameter N can be adjusted to balance between link count and performance.
3Device complexity
If network architecture is simplified to reduce complexity, then ease of implementation improves, but the ability to isolate faults and scale efficiently deteriorates
Solution Approach 1:
The N-dimensional torus architecture provides inherent fault isolation capabilities through its dimensional structure. When a fault occurs in one dimension or between specific nodes, the multi-dimensional paths allow alternative routing through other dimensions, isolating the fault's impact. The dimensional separation naturally confines faults to specific regions of the network space.
Solution Approach 2:
The regular torus structure provides universal routing rules that work for all node pairs in the network. The same dimensional routing algorithm applies regardless of network size or specific node locations, enabling efficient scaling. This universality simplifies implementation while maintaining reliability through consistent fault isolation behavior across the entire network.
4Productivity
If more communication links are added to support scaling to more nodes, then the network capacity increases, but the energy consumption and cost increase
Solution Approach 1:
The N-dimensional torus architecture enables network scaling by adding dimensions rather than simply increasing link density in a flat structure. When scaling from 2^N to 2^(N+1) nodes, the network adds a new dimension rather than exponentially increasing links per node. This dimensional scaling approach maintains constant per-node link counts while increasing overall network capacity, reducing energy consumption per node.
Solution Approach 2:
The network architecture dynamically adapts to scaling requirements by activating additional dimensions as nodes are added. The routing algorithm dynamically selects which dimensions to traverse based on source and destination coordinates, optimizing energy efficiency by choosing the shortest path through the dimensional space. This dynamic behavior allows the network to scale capacity without proportionally increasing energy consumption.
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
A data center network architecture and method for communicating data are provided. The nodes and communication links in the network are arranged according to N dimensions. Groups of four nodes are arranged initially in quad full mesh networks. In each dimension, each node has three connections to other nodes. In particular, in each dimension, the nodes can be arranged in quarters, where a node in one quarter is connected to a node in each of three other quarters. In each dimension, the maximum number of hops between a sending node and a destination node is N. Due to the dimensionality, the nodes and communication links can be easily classified into different service levels, and diagnosis of problems is facilitated.