Entry Date:
January 19, 2017

Revenue Management For Enterprise Users of Cloud Infrastructure

Principal Investigator Devavrat Shah

Co-investigator Vivek Farias

Project Start Date September 2016

Project End Date
 August 2019

The use of virtualized, cloud infrastructure has transformed the nature and scope of modern enterprise computing. Large corporations today, in place of investing in dedicated computing infrastructure, effectively rent compute infrastructure by taking advantage of both free and pay-as-you-go cloud offerings. From an economic perspective this creates a win-win situation: it lets "buyers" pay for what they use and scale their infrastructure seamlessly at low cost without investing to buy or maintain resources while creating profitable businesses for "sellers" or providers. Cost-savings are derived from effectively pooling resource needs across a number of buyers of cloud infrastructure; these cost-savings are in principle shared by providers and consumers of cloud infrastructure. However, as it stands currently buyers end up buying resources from one provider only leading to inefficiency. The goal of this project is to address the operational challenges to enable an efficient market for the enterprise cloud infrastructure. The principle investigators are committed to the mentoring of individuals from underrepresented and minority groups.

This project focuses on two parallel research thrusts: buyer-side -and seller-side problems. A key element of both these problems is the role played by uncertainty and the challenge in adopting either stochastic models (which tend to be highly unstable in this domain) or adversarial models (which tend to ignore the copious amount of historical data typically available). As such, project introduces a data-driven model that generalizes a broad class of models with a rich history in the statistics literature. This model treads the line between stochastic and adversarial modeling. In the context of this model, two important classes of problems are studied. On the buyer side, a natural version of what has become known as the k-secretary problem is studied. On the seller side, a Network Revenue Management model of demand is studied. In all of these, the guiding objective is to advance science required to develop out-of-the-box software that can be deployed by practitioners without the need for model fitting or calibration. The project will potentially lead to open-source system that allows buyers to achieve a dramatic reduction in costs taking advantage of spot markets.