Cloud configuration optimization is the procedure to determine the number and the type of instances to use when deploying an application in cloud environments, given a cost or performance objective. In the absence of a performance model for the distributed application, black-box optimization can be used to perform automatic cloud configuration. Numerous black-box optimization algorithms have been developed; however, their comparative evaluation has so far been limited to the hyper-parameter optimization setting, which differs significantly from the cloud configuration problem. In this paper, we evaluate 8 commonly used black-box optimization algorithms to determine their applicability for the cloud configuration problem. Our evaluation, using 23 different workloads, shows that in several cases Bayesian optimization with Gradient boosted regression trees performs better than methods chosen by prior work.
|Original language||English (US)|
|Number of pages||13|
|Journal||Proceedings of the VLDB Endowment|
|State||Published - Aug 2020|
Bibliographical noteKAUST Repository Item: Exported on 2021-02-16
Acknowledgements: Muhammad Bilal was supported by a fellowship from the Erasmus Mundus Joint Doctorate in Distributed Computing (EMJD-DC) program funded by the European Commission (EACEA) (FPA2012-0030). Work done in part while author was interning at KAUST. This research was supported by Fundac ̧ ̃ao para a Ciˆen-cia e a Tecnologia (FCT), under projects UIDB/50021/2020 andCMUP-ERI/TIC/0046/2014.