Utility-Directed Allocation

This paper considers the problem of allocating discrete res ources according to utility functions reported by potential recip ients and relates this abstract problem to resource allocation in a Ut ility Data Center (UDC). A simple integer program formulation, which g eneralizes well-known knapsack problems, permits a remarkab le breadth of expression while retaining clarity and analytic tractab ility. In the UDC context, this formulation allows us to incorporate fact ors such as resource scarcity, user demand, and operating costs in a u nified framework. It is equally applicable to longand short-term allocation. If applied to short-term dynamic re-allocation it all ows SLA violation penalties to be enforced or relaxed, thereby perm itting principled preemption of resources. Retrospective analys is of past allocator inputs can guide economically-optimal capacity expansion. The proposed problem formulation is suitable both for UDCs that operate exclusively within an enterprise and for those that sell access to computational resources to external customers. T he latter case involves multiple divergent interests contending for scarce resources, and this paper surveys the economic issues that a rise in such situations and relevant literature, e.g., on mechanis m design and auction theory. This paper describes the expressive power of the proposed pr oblem formulation, considers the computational requirement s of solution methods, outlines connections with economics liter ature, and compares the proposed formulation with other UDC allocatio n schemes. It also presents preliminary computational results showin g that a commercial integer-program solver can quickly find near-op timal solutions to random problem instances of reasonable size.