Group-Based Asynchronous Distributed Alternating Direction Method of Multipliers in Multicore Cluster

Abstract

The distributed alternating direction method of multipliers (ADMM) algorithm is one of the effective methods to solve the global consensus optimization problem. Considering the differences between the communication of intra-nodes and inter-nodes in multicore cluster, we propose a group-based asynchronous distributed ADMM (GAD-ADMM) algorithm: based on the traditional star topology network, the grouping layer is added. The workers are grouped according to the process allocation in nodes and model similarity of datasets, and the group local variables are used to replace the local variables to compute the global variable. The algorithm improves the communication efficiency of the system by reducing communication between nodes and accelerates the convergence speed by relaxing the global consistency constraint. Finally, the algorithm is used to solve the logistic regression problem in a multicore cluster. The experiments on the Ziqiang 4000 showed that the GAD-ADMM reduces the system time cost by 35 % compared with the AD-ADMM.