SplitPlace: Intelligent Placement of Split Neural Nets in Mobile Edge Environments

TL;DR

SplitPlace introduces an intelligent placement policy for splitting neural networks across mobile edge devices, optimizing performance despite resource constraints, and advancing distributed deep learning deployment.

Contribution

The paper proposes a novel placement algorithm, SplitPlace, for efficiently distributing neural network fragments on resource-limited edge devices.

Findings

Improves neural network deployment efficiency on edge devices.

Demonstrates scalability and performance benefits of the placement policy.

Addresses the gap in intelligent placement algorithms for split neural networks.

Abstract

In recent years, deep learning models have become ubiquitous in industry and academia alike. Modern deep neural networks can solve one of the most complex problems today, but coming with the price of massive compute and storage requirements. This makes deploying such massive neural networks challenging in the mobile edge computing paradigm, where edge nodes are resource-constrained, hence limiting the input analysis power of such frameworks. Semantic and layer-wise splitting of neural networks for distributed processing show some hope in this direction. However, there are no intelligent algorithms that place such modular splits to edge nodes for optimal performance. This work proposes a novel placement policy, SplitPlace, for the placement of such neural network split fragments on mobile edge hosts for efficient and scalable computing.