NeuroMessenger: Towards Error Tolerant Distributed Machine Learning Over Edge Networks

Abstract

Despite the evolution of distributed machine learning (ML) systems in recent years, the communication overhead induced by their data transfers remains a major issue that hampers the efficiency of such systems, especially in edge networks with poor wireless link conditions. In this paper, we propose to explore a new paradigm of error-tolerant distribute ML to mitigate the communication overhead. Unlike generic network traffic, ML data exhibits an intrinsic error-tolerant capability which helps the model yield fair performance even with errors in the data transfers. We first characterize the error tolerance capability of state-of-art distributed ML frameworks. Based on the observations, we propose NeuroMessenger, a lightweight mechanism that can be built into the cellular network stack, which can enhance and utilize the error tolerance in ML data to reduce communication overhead. NeuroMessenger does not require per-model profiling and is transparent to application layer, which simplifies the development and deployment. Our experiments on a 5G simulation framework demonstrate that NeuroMessenger reduces the end-to-end latency by up to 99% while maintaining low accuracy loss under various link conditions.