Network resources can easily become the bottleneck in a data center (DC), potentially corrupting service-level objectives. Yet most previous work on resource allocation in virtualized DCs focuses only on server resources, ignoring communication aspects. This paper thus addresses a very important research problem: how to allocate requests for virtual machines to the available servers of the DC, taking into account both server resources and bandwidth.

The authors devise a heuristic algorithm for this problem, aiming to maximize the number of requests that can be accommodated by the DC, subject to server and network capacity constraints. This is a sophisticated proprietary heuristic, which considers several candidate partitions of the communication graph of the requested virtual machines. For each candidate partition, it is checked whether it can be mapped on a set of server clusters of the DC, taking into account the capacity constraints.

A particularly interesting aspect is that the authors also experimented with integrating an existing approximation algorithm for the related minimum k-cut problem. (In the minimum k-cut problem, the aim is to find a minimum cost set of edges in an undirected graph so that after the removal of these edges, the remaining graph decomposes into k components.)

A simulation-based evaluation and comparison against a load-balancing non-network-aware resource allocator complete the paper. In the investigated scenarios, the authors’ resource allocation algorithm consistently outperforms the load-balancing non-network-aware resource allocator.

Overall, this is a must-read for people working in the field of DC management, virtual machine placement, or DC networks.