Online Model Retraining, Compression, and Instance Allocation in Edge Computing Networks

Abstract

The adoption of artificial intelligence models (e.g., DNN models) in Internet of Things has boosted computing demands in edge computing. Frequent model retraining, necessitated by concept drift, further increases resource usage, while model compression sacrifices model performance for computing efficiency. However, few works study computing instance allocation problem considering dynamic model retraining and compression, especially under varying workloads and model performance degradation. In this work, we model the problem as a joint online model retraining, compression, and instance allocation problem in edge computing networks considering model performance and instance cost. Solving the online problem is challenging since it is a non-linear binary programming problem with time-coupling instance switching cost. We first solve the online problem under fixed compression and propose an efficient online algorithm. Specifically, we first linearize the non-linear term, then regularize the time-coupling switching cost to decouple the problem, and finally use a randomization rounding method to derive the integral solution. We prove that our algorithm achieves a constant optimality gap. We then solve the online problem under flexible compression and propose a lightweight online algorithm. We extend the linearization method and decouple the problem into each time slot, and demonstrate our algorithm achieves an optimality gap depending on the time period. Simulations demonstrate that our algorithm can achieve a balance between instance cost and model performance in both fixed compression and flexible compression scenarios.