Joint VM placement and topology optimization for traffic scalability in dynamic datacenter networks

Abstract

In dynamic datacenter networks (DDNs), there are two ways to handle growing traffic: adjusting the network topology according to the traffic and placing virtual machines (VMs) to change the workload according to the topology. While previous work only focused on one of these two approaches, in this paper, we jointly optimize both virtual machine placement and topology design to achieve higher traffic scalability. We formulate this joint optimization problem to be a mixed integer linear programming (MILP) model and design an efficient heuristic based on Lagrange’s relaxation decomposition. To handle traffic dynamics, we introduce an online algorithm that can balance algorithm performance and overhead. Our extensive simulation with various network settings and traffic patterns shows that compared with randomly placing VMs in fixed datacenter networks, our algorithm can reduce up to 58.78% of the traffic in the network, and completely avoid traffic overflow in most cases. Furthermore, our online algorithm greatly reduces network cost without sacrificing too much network stability.