Resilient Mobile Distributed Computing via NDN Spark and Coded Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current mobile edge devices lack the computational power to handle intensive tasks due to limited bandwidth and unstable connectivity, making conventional centralized solutions impractical for tactical scenarios, and existing distributed computing techniques require stable network connectivity and more resources than typically available in these environments.
Innovation Solution
A resilient mobile distributed computing system combining Coded Computing (CC) and Named Data Networking (NDN) is implemented, using a Spark distributed computing network with NDN protocol and a CDC-based target recognition model to generate parity images and reconstruct missing predictions, enabling efficient processing and classification of images even in low-bandwidth and unstable conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If computationally intensive tasks are offloaded to centralized cloud, then processing power is improved, but network bandwidth and connectivity stability deteriorate
Solution Approach 1:
The patent segments the centralized cloud computing system into distributed edge computing nodes deployed across the tactical network. Each edge device becomes an independent computing node that can process tasks locally, eliminating the need to concentrate all computational power in a single centralized cloud that would require high-bandwidth network connections.
Solution Approach 2:
The patent introduces a new dimension of computation by deploying edge computing capabilities directly at the network edge (on tactical devices themselves), rather than relying solely on the traditional centralized cloud dimension. This spatial redistribution of computational resources allows processing to occur closer to the data source, reducing network dependency.
2Power
If conventional distributed computing techniques are used, then computational power is improved, but network connectivity stability and device complexity worsen
Solution Approach 1:
The patent applies preliminary action by pre-training deep neural network models on centralized cloud resources before deploying them to edge devices. This allows edge nodes to perform inference tasks with high accuracy without requiring stable network connectivity during operation, as the computational heavy lifting has already been completed in advance.
Solution Approach 2:
The patent uses copying by replicating trained model parameters and weights across multiple edge devices. Each edge node receives a copy of the pre-trained model, enabling independent local inference without requiring continuous network connection to the central server, thus improving reliability in disconnected environments.
3Productivity
If more computing resources are allocated to edge devices, then processing capability is improved, but device complexity and resource constraints worsen
Solution Approach 1:
The patent extracts the most computationally intensive components (model training and parameter optimization) from edge devices and relocates them to centralized cloud resources. Edge devices retain only the inference functionality, which requires significantly fewer computational resources, thus maintaining high processing capability while reducing device complexity and resource requirements.
4Quantity of substance
If data is processed locally on edge devices, then network bandwidth usage is reduced, but computational power and processing speed worsen
Solution Approach 1:
The patent performs preliminary model training on centralized cloud resources before deploying to edge devices. This preliminary action transfers the computationally intensive training phase to the cloud, allowing edge devices to perform only lightweight inference locally, thus reducing network bandwidth usage while maintaining adequate computational power for real-time processing.
Data Source
AI summary
A system includes: a named data networking (NDN) based Spark distributed computing network including a Spark distributed computing network including a master computer node and a plurality of slave computer nodes, and a named data networking (NDN) protocol installed on the Spark distributed computing network, and a coded distributed computing (CDC) target recognition model deployed on the NDN-based Spark distributed computing network. The NDN-based Spark distributed computing network is configured to: receive one or more batches of input images; generate a parity image from each batch of the input images; predict a label for each image of the batch of the input images; process the generated parity image; upon a label prediction of one image of the batch of the input images being unavailable, reconstruct the unavailable label prediction; and classify labels for the input images.