DynSL speeds up training by adapting model splits on edge devices
DynSL: Resource-Adaptive Model Training across Heterogeneous Edge Devices
Distributed, Parallel, and Cluster ComputingMachine LearningNetworking and Internet Architecture
Summary
Training AI models on devices like phones or sensors is hard because these devices have very different abilities. Sending all data to a central place can cause privacy problems, but training entirely on one device might be too slow or impossible. The authors created DynSL, a system that chooses which parts of the AI to run on the device and which to run on a server. It keeps changing this choice based on how fast each device and the server are at the moment. This way, training is faster without losing accuracy.
What this means in practice
- •For mobile app developers: Improve AI model training on diverse phones by dynamically shifting workloads between phone and cloud to speed up training while protecting user data.
- •For iot device teams: Enhance training efficiency on a wide range of sensors and edge devices by adapting how AI work is shared with servers during distributed learning.
Authors
Boris Radovič, Istemi Ekin Akkus, Marco Canini, Ruichuan Chen, Alice Dethise, Ivica Rimac, Veljko Pejović
Abstract
Training models over data distributed across edge devices presents a trade-off: centralizing data simplifies training but raises privacy and regulatory concerns, while full local training may exceed the devices' compute and memory capacity. Split learning addresses this challenge by partitioning the model between clients and the server, allowing part of the computation to be offloaded. However, existing approaches typically rely on static split configurations or simplified assumptions that fail to capture the heterogeneity and complex dynamics of real deployments. We present DynSL, a framework that dynamically allocates model layers between clients and the server by continuously monitoring system metrics and periodically solving an optimization problem that accounts for client and server computation and communication. When beneficial, the system reconfigures the split, enabling the offloading strategy to adapt to changing device capabilities and system conditions. Extensive experiments on a testbed of 24 heterogeneous edge devices show that DynSL reduces training time by up to $\sim\!$30% over the strongest adaptive baseline and by more than 50% over a suboptimal uniform split, with no loss -- and sometimes a gain -- in accuracy.