Placement of parameter server in wide area network topology for geo-distributed machine learning

Machine learning (ML) is extensively used in a wide range of real-world applications that require data all around world to pursue high accuracy global model. Unfortunately, it impossible transmit gathered raw central center for training due privacy, sovereignty and communication cost. This brings the idea geo-distributed machine (Geo-DML), which completes ML model across multiple centers with bottleneck cost over limited area networks (WAN) bandwidth. In this paper, we study on problem parameter server (PS) placement PS architecture efficiency Geo-DML. Our optimization aims select an appropriate as algorithm based We prove NP-hard. Further, develop approximation solve using randomized rounding method. order validate performance our proposed algorithm, conduct large-scale simulations, simulation results two typical carrier network topologies show can reduce up 61.78% B4 topology 21.78% Internet2 topology.