Multi-Container Application Migration with Load Balanced and Adaptive Parallel TCP

Abstract

Application migration in Wide Area Network (WAN) is needed in many scenarios: disaster recovery and service hand-off in edge cloud. Modern distributed applications are virtualized with multiple virtual machines or containers. This paper focuses on parallel multi-container migration in WAN environments by utilizing multiple TCP connections over a single direct path (a.k.a parallel TCP). Our application migration middleware framework utilizes a feedback controller to determine a proper number of parallel container migration (i.e., parallel window) based on changing network bandwidth. Then a middleware’s scheduler selects migration requests for the parallel window to load balance multiple pairs of hosts. The goal of our migration is to achieve the best possible balance between optimizing the total migration time and average individual migration time. This differs from most existing live migration works that attempt to optimize mainly the down time. Our proposed framework is generalized and not restricted to any particular virtualization technology implementation. For performance evaluation, we conducted a WAN-emulated experiment in static and dynamic network settings. The performance evaluation results show that the total migration time using our feedback controller is less than that of the sequential migration by 32.7% in the static network, and 43.9% in the dynamic network. Moreover, while achieving total migration time comparable to that of the best fixed parallel migration window, our approach can reduce the average individual migration time by 62.7% in the dynamic network setting.