Computing Node Interconnects via Parallel High-Radix Switch Links
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing high-speed data communication networks face challenges in efficiently interconnecting large numbers of computing nodes due to the tradeoff between network device radix and bandwidth, limiting the flattening of network topology and increasing latency and power consumption.
Innovation Solution
Implementing a system with high-radix network switches that utilize parallel transmission of low-bandwidth data streams without serialization or deserialization, using optical transceivers for communication between computing nodes and switches via free-space or optical cables, enabling efficient interconnection of computing nodes in datacenters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If SerDes is used to convert multiple parallel data streams into a single high-bandwidth serial channel, then transmission bandwidth is improved, but network device complexity and latency increase
Solution Approach 1:
The patent divides the network into multiple parallel low-bandwidth channels instead of using a single high-bandwidth serial channel. Each computing node has multiple network interface cards (NICs) that transmit data streams in parallel through separate network switches, avoiding the need for SerDes serialization/deserialization at network devices and reducing device complexity.
2Productivity
If SerDes is used to transmit high-bandwidth serial channels, then transmission efficiency is improved, but latency increases
Solution Approach 1:
Data transmission is segmented into multiple parallel low-bandwidth channels that can be transmitted simultaneously without serialization overhead. This parallel transmission approach eliminates the time required for SerDes conversion processes, reducing latency while maintaining overall transmission efficiency through concurrent data flow.
3Speed
If network topology is flattened to reduce latency, then speed is improved, but the tradeoff between radix and bandwidth limits the degree of flattening
Solution Approach 1:
The network is segmented into multiple parallel paths with lower individual bandwidth requirements, enabling network switches to have higher radix (more ports) without requiring each port to support extremely high bandwidth. This allows for greater network topology flattening with multiple layers of switches, reducing latency through shorter transmission paths.
4Loss of time
If high-radix network switches are used to flatten network topology, then latency is reduced, but bandwidth requirements per channel increase
Solution Approach 1:
The total bandwidth requirement is segmented across multiple parallel channels with lower individual bandwidth demands. This allows network switches to be designed with higher radix (more connection points) while maintaining moderate bandwidth per port, achieving network topology flattening without requiring extremely high bandwidth per channel.
Data Source
AI summary
A system for efficiently interconnecting computing nodes can include a plurality of computing nodes and a plurality of network switches coupled in parallel to the plurality of computing nodes. The system can also include a plurality of node interfaces. Each computing node among the plurality of computing nodes can include at least one node interface for each network switch among the plurality of network switches. The plurality of node interfaces corresponding to a computing node can be configured to send data to another computing node via the plurality of network switches. The system can also include a plurality of switch interfaces. Each network switch among the plurality of network switches can include at least one switch interface for each computing node among the plurality of computing nodes. A switch interface corresponding to the computing node can be coupled to a node interface corresponding to the computing node.