Energy-aware scheduling in distributed computing systems

Iturriaga, Santiago

Supervisor(es): Nesmachnow, Sergio - Dorronsoro, Bernabé

Resumen:

Distributed computing systems, such as data centers, are key for supporting modern computing demands. However, the energy consumption of data centers has become a major concern over the last decade. Worldwide energy consumption in 2012 was estimated to be around 270 TWh, and grim forecasts predict it will quadruple by 2030. Maximizing energy efficiency while also maximizing computing efficiency is a major challenge for modern data centers. This work addresses this challenge by scheduling the operation of modern data centers, considering a multi-objective approach for simultaneously optimizing both efficiency objectives. Multiple data center scenarios are studied, such as scheduling a single data center and scheduling a federation of several geographically-distributed data centers. Mathematical models are formulated for each scenario, considering the modeling of their most relevant components such as computing resources, computing workload, cooling system, networking, and green energy generators, among others. A set of accurate heuristic and metaheuristic algorithms are designed for addressing the scheduling problem. These scheduling algorithms are comprehensively studied, and compared with each other, using statistical tools to evaluate their efficacy when addressing realistic workloads and scenarios. Experimental results show the designed scheduling algorithms are able to significantly increase the energy efficiency of data centers when compared to traditional scheduling methods, while providing a diverse set of trade-off solutions regarding the computing efficiency of the data center. These results confirm the effectiveness of the proposed algorithmic approaches for data center infrastructures.


Los sistemas informáticos distribuidos, como los centros de datos, son clave para satisfacer la demanda informática moderna. Sin embargo, su consumo de energético se ha convertido en una gran preocupación. Se estima que mundialmente su consumo energético rondó los 270 TWh en el año 2012, y algunos prevén que este consumo se cuadruplicará para el año 2030. Maximizar simultáneamente la eficiencia energética y computacional de los centros de datos es un desafío crítico. Esta tesis aborda dicho desafío mediante la planificación de la operativa del centro de datos considerando un enfoque multiobjetivo para optimizar simultáneamente ambos objetivos de eficiencia. En esta tesis se estudian múltiples variantes del problema, desde la planificación de un único centro de datos hasta la de una federación de múltiples centros de datos geográficmentea distribuidos. Para esto, se formulan modelos matemáticos para cada variante del problema, modelado sus componentes más relevantes, como: recursos computacionales, carga de trabajo, refrigeración, redes, energía verde, etc. Para resolver el problema de planificación planteado, se diseñan un conjunto de algoritmos heurísticos y metaheurísticos. Estos son estudiados exhaustivamente y su eficiencia es evaluada utilizando una batería de herramientas estadísticas. Los resultados experimentales muestran que los algoritmos de planificación diseñados son capaces de aumentar significativamente la eficiencia energética de un centros de datos en comparación con métodos tradicionales planificación. A su vez, los métodos propuestos proporcionan un conjunto diverso de soluciones con diferente nivel de compromiso respecto a la eficiencia computacional del centro de datos. Estos resultados confirman la eficacia del enfoque algorítmico propuesto.


Detalles Bibliográficos
2017
Data centers
Energy efficiency
Job scheduling
Centro de datos
Eficiencia energética
Planificación de tareas
Inglés
Universidad de la República
COLIBRI
https://hdl.handle.net/20.500.12008/21701
Acceso abierto
Licencia Creative Commons Atribución – No Comercial – Sin Derivadas (CC-BY-NC-ND)
Resumen:
Sumario:Distributed computing systems, such as data centers, are key for supporting modern computing demands. However, the energy consumption of data centers has become a major concern over the last decade. Worldwide energy consumption in 2012 was estimated to be around 270 TWh, and grim forecasts predict it will quadruple by 2030. Maximizing energy efficiency while also maximizing computing efficiency is a major challenge for modern data centers. This work addresses this challenge by scheduling the operation of modern data centers, considering a multi-objective approach for simultaneously optimizing both efficiency objectives. Multiple data center scenarios are studied, such as scheduling a single data center and scheduling a federation of several geographically-distributed data centers. Mathematical models are formulated for each scenario, considering the modeling of their most relevant components such as computing resources, computing workload, cooling system, networking, and green energy generators, among others. A set of accurate heuristic and metaheuristic algorithms are designed for addressing the scheduling problem. These scheduling algorithms are comprehensively studied, and compared with each other, using statistical tools to evaluate their efficacy when addressing realistic workloads and scenarios. Experimental results show the designed scheduling algorithms are able to significantly increase the energy efficiency of data centers when compared to traditional scheduling methods, while providing a diverse set of trade-off solutions regarding the computing efficiency of the data center. These results confirm the effectiveness of the proposed algorithmic approaches for data center infrastructures.