Optical Reshufflers for Single-Layer Data Center Network Scalability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Structured network architectures in data centers face issues such as increased latency due to multiple hops, difficulty in horizontal scaling, underutilization of resources, and susceptibility to cluster-packing problems, while purely random networks are challenging to construct, manage, and visualize.
Innovation Solution
A structured rearranged network architecture using optical reshufflers to randomly interconnect switches in a single layer, with structural constraints that prioritize diversity and reachability, allowing for higher throughput, lower latency, and easier scalability, while maintaining manageability and visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If structured network architectures (e.g., Fat Tree, Clos) are used to provide high bisectional bandwidth and ease of routing, then network reliability and ease of operation are improved, but latency increases due to multiple hops and device complexity increases with additional layers
Solution Approach 1:
The patent segments the network into two distinct functional layers: an unstructured access layer for random connectivity and a structured core layer for reliable routing. This segmentation allows traffic to benefit from randomization at the access point while maintaining structured paths through the core, thereby reducing overall latency while preserving reliability.
Solution Approach 2:
The patent introduces an intermediary structured core layer that mediates between the unstructured access layer and the destination. This intermediary provides deterministic routing paths and reliability guarantees while the access layer provides low-latency random connectivity, resolving the contradiction between reliability and latency.
2Ease of manufacture
If structured network architectures are deployed in discrete implementation sizes to simplify construction, then ease of manufacture is improved, but adaptability decreases as higher layer ports may go unused in underfilled networks
Solution Approach 1:
The patent makes the network dynamic by allowing the unstructured access layer to adapt its connectivity patterns based on actual traffic demands. The structured core layer maintains its formal structure for reliability, while the access layer dynamically adjusts to fill utilization, enabling the network to adapt to varying loads without requiring discrete size implementations.
3Device complexity
If hardware devices are port limited by available pins, then device complexity is reduced, but horizontal scaling capability decreases as bandwidth increases require more ports
Solution Approach 1:
The patent transitions from horizontal scaling within a single layer to vertical scaling across two layers. By stacking multiple unstructured access layers on top of a structured core layer, the network achieves horizontal scaling capability without increasing the port complexity of individual devices. Each device remains relatively simple while the multi-layer architecture provides extensive scaling capability.
4Productivity
If a purely random network approach is used to overcome structured network limitations, then adaptability and throughput are improved, but device complexity increases for construction and management
Solution Approach 1:
The patent applies local quality by making each layer specialized: the unstructured access layer handles random connectivity and high throughput locally, while the structured core layer handles routing and management deterministically. This division of functional qualities allows the system to achieve high throughput through randomization without imposing the associated management complexity on the entire network.
Data Source
AI summary
A data center network utilizing a single layer architecture includes a plurality of switches each with a plurality of ports including a first set of ports of the plurality of ports including network facing ports and a second set of ports of the plurality of ports including server facing ports; and a plurality of optical reshufflers configured to randomly interconnect the plurality of switches via the network facing ports of each in a single layer, unstructured network based on a plurality of structural constraints. The value of a number of the network facing ports is equal or greater than a number of the server facing ports and wherein each of the plurality of switches is a switch with attached servers. The plurality of optical reshufflers can be spatially partitioned across a layer with each optical reshuffler restricted to internal connectivity.