Near-optimal virtual machine placement with product traffic pattern in data centers

This paper focuses on optimizing the virtual machine (VM) placement to reduce the bandwidth usage in modern virtualization based data centers. Existing solutions require all the traffic rates between any pair of VMs to be known beforehand, which may incur significant overhead or even be impossible. To reduce such overhead, we adopt the product traffic pattern model to characterize the traffic rates. In this model, each VM is associated with an activity level, and the normalized traffic rate between each pair of VMs is approximated by the product of their activity levels. Specifically, we consider the VM placement problem with product traffic in data centers that follow the Clique or VL2 architecture. We then present an optimal solution for the special case where all the physical machines (PMs) connect to the same switch. By extending the intrinsic idea of such solution, we further develop a simple yet efficient algorithm for the general case. We prove its optimality in a common case where the PMs are homogeneous, and also derive its approximation ratio in the general case. Finally, we demonstrate via simulations that the performance of our algorithm is near-optimal under practical settings.