Dolphin : Runtime Optimization for Distributed Machine Learning

Large-scale machine learning (ML) systems are becoming widely used. Typically, these ML systems run on fixed resources, but it is difficult to find their optimal configurations (e.g., how many nodes to use, how to distribute data) since they depend on multiple factors such as hardware environments, ML algorithms, input datasets, etc. Furthermore, optimal configurations often can change over time due to fluctuating cluster resources and changing ML algorithm patterns. In this paper, we present Dolphin, an elastic machine learning framework that addresses the configuration problem at runtime. Dolphin solves a cost-based optimization problem to find an optimal configuration and reconfigures the system dynamically at runtime. Dolphin introduces a new distributed memory abstraction to change resource and data configurations based on the optimizer plan transparently and efficiently.