No Regret in Cloud Resources Reservation with Violation Guarantees

Abstract

This paper addresses a fundamental challenge in cloud computing, that of learning an economical yet robust reservation, i.e. reserve just enough resources to avoid both violations and expensive over provisioning. Prediction tools are often inadequate due to observed high variability in CPU and memory workload. We propose a novel model-free approach that has its root in online learning. Specifically, we allow the workload profile to be engineered by an adversary who aims to harm our decisions, and we investigate a class of policies that aim to minimize regret (minimize losses with respect to a baseline static policy that knows the workload sample path). Then we propose a combination of the Lyapunov optimization theory [1] and a linear prediction of the future based on the recent past, used in learning and online optimization problems, see [2]. This enables us to come up with a no regret policy, i.e., a policy whose cost difference to the benchmark and violation constraint residual both grow sublinearly in time, and hence become amortized over the horizon. Our policy has then “no regret and eventually learns the minimum cost reservation subject to a time-average constraint for violations.