Joint data placement and flow control for cost-efficient data center networks

We study the problem of cost-efficient operation of data center networks used to deliver heterogenous online services. We split their aggregate cost into server-load-related and link-load-related segments. Thus, we formulate the problem of interest as that of joint data placement and flow control and use mixed integer-linear programming to compute the optimal solution. The high complexity of the latter motivated us to design two additional sets of strategies, based on data coding and heuristics, respectively. In our simulation experiments, carried out based on actual data center information, network topology and link cost, and electricity prices, we examine the advantages of data coding, in particular in the context of multicast, and the impact of different factors such as the network topology and service popularity, on the total cost incured by all strategies we consider. We show that network coding with multicast provides cost savings on the order of 30-80%, depending on the specific context under consideration, relative to the other optimization strategies and heuristic methods that we examine.