Automatic parameter tuning for Class-Based Virtual Machine Placement in cloud infrastructures

A critical task in the management of Infrastructure as a Service cloud data centers is the placement of Virtual Machines (VMs) over the infrastructure of physical nodes. However, as the size of data centers grows, finding optimal VM placement solutions becomes challenging. The typical approach is to rely on heuristics that improve VM placement scalability by (partially) discarding information about the VM behavior. An alternative approach providing encouraging results, namely Class-Based Placement (CBP), has been proposed recently. CBP considers VMs divided in classes with similar behavior in terms of resource usage. This technique can obtain high quality placement because it considers a detailed model of VM behavior on a per-class base. At the same time, scalability is achieved by considering a small-scale VM placement problem that is replicated as a building block for the whole data center. However, a critical parameter of CBP technique is the number (and size) of building blocks to consider. Many small building blocks may reduce the overall VM placement solution quality due to fragmentation of the physical node resources over blocks. On the other hand, few large building blocks may become computationally expensive to handle and may be unsolvable due to the problem complexity. This paper addresses this problem analyzing the impact of block size on the performance of the VM class-based placement. Furthermore, we propose an algorithm to estimate the best number of blocks. Our proposal is validated through experimental results based on a real cloud computing data center.