Virtual machine scheduling for energy and sla optimisation based in metaheuristics and forecasting techniques in cloud computing
- Autores: Sergi Vila Almenara
- Directores de la Tesis: Fernando Guirado (codir. tes.), J.L. Lérida (codir. tes.)
- Lectura: En la Universitat de Lleida ( España ) en 2022
- Idioma: español
- Tribunal Calificador de la Tesis: José Luis Bosque Orero (presid.), Concepció Roig Mateu (secret.), João Pedro Barreto Laurindo (voc.)
- Programa de doctorado: Programa de Doctorado en Ingeniería y Tecnologías de la Información por la Universidad de Lleida
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Ever since computing and the Internet became part of our society, large datacenters have been necessary to serve customers, businesses, governments and the scientific community. Machine virtualisation has made it possible to encapsulate multiple instances of operating systems on a single physical device, allowing for lower power consumption in resource sharing. Because these resource requirements vary over time, corrections in the allocation of virtual machines to other physical machines must be applied to guarantee Service Level Objectives (SLOs).
The main purpose of this thesis is to contribute to Cloud environments by improving energy consumption and Quality of Service (QoS), by means of the Service Level Agreement (SLA), through techniques that allow the correct decisions to be executed at the correct moment. Resource planning allows the needs and demands of each physical and virtual machine to be adapted to improve the overall system performance. By performing optimal task assignations and Virtual Machine migrations, datacenters become more stable and efficient, requiring less supervision and corrections, guaranteeing the indispensable number of active physical machines running properly and avoiding the performing of redundant actions.
Two types of environments were dealt with during this thesis. Firstly, static environments were the focus, where the Bin Packing Problem was solved to optimise the energy and makespan proposing a multi-objective genetic algorithm (BLEMO). The next focus was on dynamic environments, in which different proposals were developed.
Initially, a virtual machine selection algorithm (WPSP) was proposed to reduce the number of migrations and SLA violations. The proposal included a VM to host an allocation algorithm (MDG) in order to reduce the total inter-host communications, and so optimising the network bandwidth. Next, the use of Bollinger Bands (BB) was combined with the NeuralProphet framework (BB+NeoPro) to generate longterm predictions. These enable consolidation of the VM to host allocations, reducing the number of migrations, the energy consumption and SLA Violations. Furthermore, an improved version of BB+NeoPro was proposed combining WPSP, BB and Facebook Prophet (WBF). This proposal was tested with different Cloud workloads from several providers in order to validate its robustness.
The analysis of the results obtained showed how all the evaluated metrics obtained substantial benefits thanks to the multi-purpose exploration of the search space in the case of static environments, and the careful and adapted observation of the available data in dynamic environments that allowed better decisions to be taken.
