Lynceus: Long-Sighted, Budget-Aware Online Tuning of Cloud Applications

Over the past years, modern cloud providers have widely incremented the heterogeneity of the products they offer, be it virtual machines or remote storage systems. With such a huge variety of machines to choose from, several configurations – sets of machines – can be formed. These configurations may yield similar performances with identical costs. Therefore, users have a hard time finding the best one for deploying their specific workloads. This paper presents Lynceus, a long-sighted, budgetaware system for self-tuning of cloud applications, which uses a roll-out technique to search the space of possible configurations for the next best configuration to try, without the overhead of a brute force search. We present numerical experiments demonstrating the cost savings achieved by our system when compared to state-of-the-art systems that neglect the cost of the exploration phase.