|PINCIROLI RICCARDO||Cycle: XXX |
Section: Computer Science and Engineering
Tutor: PERNICI BARBARA Major Research topic
:Energy efficiency in large data-centers using performance evaluation techniques
Advisor: GRIBAUDO MARCOAbstract:
Nowadays, the amount of available computing resources is limitless and the energy consumption in IT is assuming an increasing importance due to the huge number of compute-intensive applications used in business and scientific fields. In 2014 the IT infrastructures in United States consumed about 70 billion kWh and this data has been estimated to grow to 73 billion kWh by 2020. Besides the significant amounts of money (e.g., 7.4 billion dollars in 2011), large energy consumption values also come with enormous volume of greenhouse gases released into the atmosphere.
While several efforts have been done to improve the data-centers’ energy efficiency, their energy consumption must be further reduced, for example, taking into consideration the workload heterogeneity. In this thesis, new techniques to evaluate and improve the energy efficiency of data-centers are proposed. To this end, performance modeling approaches such as queuing networks, Petri nets and stochastic processes are applied, and analytical and discrete event simulation techniques are used.
To investigate the data-centers energy consumption problem, four different paths are followed: i) a new power model is proposed for better characterize current systems’ architecture; ii) a new energy metric is introduced to account for data-centers’ workload heterogeneity; iii) a new framework is adopted to study Big data applications; iv) epistemic uncertainty is propagated into a model to improve the accuracy of its output measures.
The power model proposed in this thesis accounts for both dynamic voltage/frequency scaling and simultaneous multi-threading, two widespread energy saving techniques used in multi-core architectures. The energy per time-unit of execution is the new energy-performance trade-off metric that also considers the heterogeneity of the workload. Pool depletion systems is the framework proposed to analyze all the applications that generate a large number of tasks which must be executed by one or more subsystems with limited capacity (e.g., Big Data applications), and investigate new scheduling strategies to decrease their execution time. Finally, since stochastic models are fundamental in this thesis, epistemic uncertainty propagation is studied. It allows the modeler to take into account the impact of uncertain input parameters on the output measures of the model.