A Separation Principle for Optimal IaaS Cloud Computing Distribution

Due to the raising importance of cloud computing and infrastructure as a service (IaaS), the first markets for the exchange of computational power over the internet are being implemented. As of today, bandwidth constraints are not explicitly embedded in these market mechanisms. In this paper, the problem of optimal allocation of the computing power and of the corresponding data flows, according to bandwidth and computing capacity constraints, is modeled as a bilevel optimization program. It is shown that this program, which is generally non convex and hard to solve, has the same optimal solution of its convex relaxation. This allows to state a fundamental separation result, showing how the congestion control protocols employed in the network do not affect the optimal allocation problem, and allows to compute the shadow prices of the available computational resources.