HiveMind: Towards Cellular Native Machine Learning Model Splitting
Song Wang, Xinyu Zhang, Hiromasa Uchiyama, Hiroki Matsuda
Abstract
The increasing processing load of today’s mobile machine learning (ML) application challenges the stringent computation budget of mobile user equipment (UE). With the wide deployment of 5G edge-cloud, a new ML offloading scheme called split ML is provisioned to enable computation-intensive mobile ML applications by splitting an ML model across mobile UE, edge, and cloud. However, the complex split assignment problems pose new challenges for split ML system design. In this paper, we introduce HiveMind, the first practical multi-split ML system tailored for 5G cellular networks. HiveMind reformulates the complicated multi-split problem to a min-cost graph search and optimizes the distributed algorithm to drastically reduce the signaling overhead. Benefit from its low overhead property, HiveMind makes the optimal split decision on multiple computing nodes in real-time and adapts the split decisions to the instantaneous network dynamics. HiveMind also incorporates a multi-objective mechanism that accommodates heterogeneous objectives for a single ML task. HiveMind adapts to a wide range of ML frameworks, including non-linear models like Recurrent Neural Network (RNN), Federated Learning (FL), and Multi-agent Reinforcement Learning (MARL). We evaluate HiveMind on 5G MEC network simulators with realistic traffic patterns and real-life MEC computation/communication profiles. Our experiments demonstrate that HiveMind achieves the optimal efficiency comparing to state-of-art split ML designs.