Machine learning (ML) is extensively used in a widerange of real-world applications that require data all aroundworld to pursue high accuracy of a global model. Unfortunately,it is impossible to transmit all gathered raw data to a centraldata center for training due to data privacy, data sovereigntyand high communication cost. This brings the idea of geo-distributed machine learning (Geo-DML), which completes thetraining of the global ML model across multiple data centerswith the bottleneck of high communication cost over the limitedwide area networks (WAN) bandwidth. In this paper, we studyon the problem of parameter server (PS) placement in PSarchitecture for communication efficiency of Geo-DML. Ouroptimization aims to select an appropriate data center as thePS for global training algorithm based on the communicationcost. We prove the PS placement problem is NP-hard. Further,we develop an approximation algorithm to solve the problemusing the randomized rounding method. In order to validate theperformance of our proposed algorithm, we conduct large-scalesimulations, and the simulation results on two typical carriernetwork topologies show that our proposed algorithm can reducethe communication cost up to 61.78% over B4 topology and21.78% over Internet2 network topology.