Neural Network Routing in Network Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional network devices rely on static rules for routing data packets, which are inadequate in dynamically determining optimal routes, leading to inefficiencies in routing times and network performance.
Innovation Solution
A device utilizing a trained neural network, specifically a Boltzmann Machine, in conjunction with programmable hardware like FPGAs and the P4 data-plane programming language, to parse input headers, determine forwarding recommendations, and process packets based on available resources, thereby improving routing decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If static rules are used for routing data packets, then device complexity is reduced and ease of operation is improved, but routing performance deteriorates and transit time increases
Solution Approach 1:
The patent applies dynamics by transitioning from static routing rules to dynamic routing decisions using a trained neural network. The neural network continuously learns from network conditions and adapts routing decisions in real-time, allowing the system to respond to changing network states while maintaining manageable complexity through the fixed architecture of the neural network model.
Solution Approach 2:
The patent replaces the mechanical system of static rule-based routing with a neural network-based intelligent system. The neural network processes packet headers and network state information to generate optimized routing decisions, substituting traditional deterministic routing logic with a learned probabilistic model that can handle complex routing scenarios more effectively.
2Ease of operation
If static rules are used for routing data packets, then ease of operation is improved, but transit time increases
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network offline using historical network data and routing patterns. This pre-training phase prepares the model to make rapid routing decisions during operation, eliminating the need for complex real-time calculations while maintaining high performance. The network device can then directly apply the learned model to incoming packets without extensive processing.
3Productivity
If a trained neural network is used for routing decisions, then routing performance is improved and transit time is reduced, but device complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the neural network implementation into distinct functional components: input processing module for parsing packet headers, neural network inference engine for generating routing decisions, and output module for forwarding packets. This modular architecture allows each component to be optimized independently and simplifies deployment on FPGA hardware by distributing different parts of the network across available resources.
Solution Approach 2:
The patent applies universality by designing a neural network architecture that can handle multiple routing scenarios and packet types through a single unified model. The neural network is trained on diverse network conditions and traffic patterns, enabling it to generalize across different routing situations without requiring separate specialized modules for each scenario, thereby reducing overall system complexity.
4Loss of time
If a trained neural network is used for routing decisions, then transit time is reduced, but device complexity increases
Solution Approach 1:
The patent applies copying by implementing the neural network model in hardware on the FPGA, creating a dedicated copy of the inference engine that processes routing decisions in parallel with network traffic flow. This hardware implementation eliminates the need for software-based processing, significantly reducing transit time while containing complexity within the fixed hardware architecture of the FPGA device.
Data Source
AI summary
A device includes a transceiver and processing circuitry. The transceiver is configured to receive an input packet having an input header and forward an output packet having an output header. The processing circuitry is configured to parse the input header, determine recommendations for forwarding a payload of the input packet using a trained neural network and based on the parsed input header, and process the input packet and generate the output packet with the output header based on the recommendations and available resources.