Dynamic resource allocation using performance forecasting

Paulo Moura, Fabio Kon, Spyros Voulgaris, Maarten Van Steen

    Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

    Abstract

    To benefit from the performance gains and cost savings enabled by elasticity in cloud IaaS environments, effective automated mechanisms for scaling are essential. This automation requires monitoring system status and defining criteria to trigger allocation and deallocation of resources. While these criteria are usually based upon resource utilization, this may be inadequate due to the impossibility of identifying the actual amount of required resources. Consequently, most systems update resource allocation in small increments, e.g., one VM at a time, which may negatively affect performance and cost. In this paper, we propose a novel approach in which the system monitors workload, instead of utilization, and, by means of a scalability model, it makes predictions of resource demand and updates the allocated resources accordingly. We provide an implementation of this approach and describe experimental results that show its effectiveness.

    Original languageEnglish
    Title of host publication2016 International Conference on High Performance Computing and Simulation, HPCS 2016
    EditorsVesna Zeljkovic, Waleed W. Smari
    PublisherIEEE
    Pages18-25
    Number of pages8
    ISBN (Electronic)9781509020881
    DOIs
    Publication statusPublished - 13 Sep 2016
    Event14th International Conference on High Performance Computing and Simulation, HPCS 2016 - Innsbruck, Austria
    Duration: 18 Jul 201622 Jul 2016
    Conference number: 14
    http://hpcs2016.cisedu.info/

    Conference

    Conference14th International Conference on High Performance Computing and Simulation, HPCS 2016
    Abbreviated titleHPCS 2016
    CountryAustria
    CityInnsbruck
    Period18/07/1622/07/16
    Internet address

    Fingerprint

    Dive into the research topics of 'Dynamic resource allocation using performance forecasting'. Together they form a unique fingerprint.

    Cite this