Lotus: A New Topology for Large-scale Distributed Machine Learning

Machine learning is at the heart of many services provided by data centers. To improve the performance of machine learning, several parameter (gradient) synchronization methods have been proposed in the literature. These synchronization algorithms have different communication characteristics and accordingly place different demands on the network architecture. However, traditional data-center networks cannot easily meet these demands. Therefore, we analyze the communication profiles associated with several common synchronization algorithms and propose a machine learning-oriented network architecture to match their characteristics. The proposed design, named Lotus, because it looks like a lotus flower, is a hybrid optical/electrical architecture based on arrayed waveguide grating routers (AWGRs). In Lotus, a complete bipartite graph is used within the group to improve bisection bandwidth and scalability. Each pair of groups is connected by an optical link, and AWGRs between adjacent groups enhance path diversity and network reliability. We also present an efficient routing algorithm to make full use of the path diversity of Lotus, which leads to a further increase in network performance. Simulation results show that the network performance of Lotus is better than Dragonfly and 3D-Torus under realistic traffic patterns for different synchronization algorithms.