Multi-Fabric Interconnection for Scalable Network Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing network designs for scalable switched networks face limitations in reliability and performance, particularly with crossbar switches employing wormhole routing, and fail to optimize both simultaneously, especially in large-scale systems with multiple incomplete fabrics.
Innovation Solution
A multi-fabric interconnection system where each node interfaces with more than two fabrics, with each fabric being incomplete, ensuring every pair of nodes connects through at least one common fabric, using balanced incomplete block designs to determine optimal fabric sizes and numbers, allowing for scalable, high-performance, and fault-tolerant connections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional two-port ServerNet networks are used with colored ports connected to two independent groups of crossbar switches, then fault tolerance is provided, but device complexity and hardware resources increase significantly as networks scale
Solution Approach 1:
The patent applies universality by allowing nodes to have ports of multiple types (X and Y ports) that can connect to different fabrics, enabling a single node to participate in multiple interconnection domains. This multi-functional capability allows the network to achieve fault tolerance without requiring complete independence of connection groups, thereby reducing overall hardware complexity while maintaining reliability through diverse routing paths.
Solution Approach 2:
The patent segments the interconnection network into multiple independent fabrics, where each fabric is a complete interconnection group that can operate autonomously. By dividing the network into separate fabrics rather than using a single large interconnection structure, the system achieves fault isolation (improving reliability) while each fabric maintains manageable size (controlling device complexity). Nodes can failover between fabrics without affecting the entire network.
2Ease of operation
If ring topologies are used for wormhole-routed switched networks, then connectivity is provided, but hop count becomes unacceptably large as the number of nodes is scaled
Solution Approach 1:
The patent transitions from a single-dimensional ring topology to a multi-dimensional fabric-based architecture where nodes can communicate through multiple independent paths across different fabrics. This dimensional expansion allows packets to choose optimal routes through the interconnection network, reducing the effective hop count between nodes while maintaining comprehensive connectivity. The fabric structure provides direct routing capabilities that eliminate the sequential traversal required in ring topologies.
3Device complexity
If bus-oriented interconnection topologies are used for single-hop communication, then simplicity is achieved, but scalability is limited because a single-hop interconnection between a large number of nodes is impossible with crossbar switches having a limited number of ports
Solution Approach 1:
The patent divides the large-scale interconnection problem into multiple smaller fabric segments, where each fabric can handle a manageable number of nodes through single-hop communication. By segmenting the overall network into multiple fabrics rather than attempting a single large bus or crossbar, the system maintains the simplicity of direct communication within each fabric while achieving system-level scalability through the collection of fabrics. This segmentation allows the network to grow by adding more fabrics rather than increasing the complexity of individual interconnection structures.
4Productivity
If non-bus-oriented single-hop interconnections are used to connect nodes directly, then connection efficiency is improved, but the design is limited to end nodes having a large number of ports and fabrics having zero switches
Solution Approach 1:
The patent creates a universal architecture where fabrics can serve multiple functions: they provide direct single-hop communication for nodes within the same fabric while also enabling switched interconnection through crossbar switches for nodes in different fabrics. Nodes are designed with multi-functionality, possessing both X and Y ports that can participate in different fabric types. This universality allows the system to achieve connection efficiency through single-hop paths when available while maintaining applicability to switched interconnect architectures through the fabric switching mechanism.
Data Source
AI summary
Interconnect networks are described that allow nodes having more than two ports to be interconnected. More particularly, each node interfaces with multiple and more than two, fabrics. Also, all fabrics are incomplete in that not every node interfaces with every fabric, and no fabric includes all the nodes, yet every pair of nodes appears together in at least one fabric. Nodes are used that appear together in a fabric as a class of nodes that exhibit similar interconnection properties. The present invention allows for scalable, high-performance and reliable interconnection of large numbers of end nodes while satisfying constraints on architecture of end nodes and networking equipment. Bounds for the number of fabrics and fabric size are disclosed for designing an optimized interconnection network.