The PhD thesis defense of Dimitra Politaki which will take place on Tuesday July 16th, 2019 at 2pm in the Euler Violet room, Inria Sophia Antipolis.
Title: “On Modeling Green Data Center Clusters”
- Georges Da Costa, IRIT, France
- Matteo Sereno, University of Turin, Italie
- Vincenso Mascunzo, IMDEA, Espagne
- Michel Riveill, Université de Côte d’Azur, France
- Sara Alouf, Inria, France
- Fabien Hermenier, Nutanix
Data center clusters energy consumption is rapidly increasing making them the fastest-growing consumers of electricity worldwide. Renewable electricity sources and especially solar energy as a clean and abundant energy can be used, in many locations, to cover their electricity needs and make them “green” namely fed by photovoltaics. This potential can be explored by predicting solar irradiance and assessing the capacity provision for data center clusters. In this thesis we develop stochastic models for solar energy; one at the surface of the Earth and a second one which models the photovoltaic output current. We then compare them to the state of the art on-off model and validate them against real data. We conclude that the solar irradiance model can better capture the multiscales correlations and is suitable for small scale cases. We then propose a new job life-cycle of a complex and real cluster system and a model for data center clusters that supports batch job submissions and considers both impatient and persistent customer behavior. To understand the essential computer cluster characteristics, we analyze in detail two different workload type traces; the first one is the published complex Google trace and the second, simpler one, which serves scientific purposes, is from the Nef cluster located at the research center Inria Sophia Antipolis. We then implement the marmoteCore-Q, a tool for the simulation of a family of queueing models based on our multi-server model for data center clusters with abandonments and resubmissions.
Keywords: Renewable energy, solar power, semi-Markov process, queueing theory, cloud computing, data center cluster, workload characterisation, Google cluster, Nef cluster.