Full text
| |
| Author(s): |
Thiago Augusto Lopes Genez
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2017-12-13 |
| Examining board members: |
Edmundo Roberto Mauro Madeira;
Carlos Alberto Kamienski;
Paulo Romero Martins Maciel;
Islene Calciolari Garcia;
Leandro Aparecido Villas
|
| Advisor: | Luiz Fernando Bittencourt; Edmundo Roberto Mauro Madeira |
| Abstract | |
To derive efficient schedules for the tasks of scientific applications modelled as workflows, schedulers need information on the application demands as well as on the resource availability, especially those regarding the available bandwidth. However, the lack of precision of bandwidth estimates provided by monitoring/measurement tools should be considered by the scheduler to achieve near-optimal schedules. Uncertainties of available bandwidth can be a result of imprecise measurement and monitoring network tools and/or their incapacity of estimating in advance the real value of the available bandwidth expected for the application during the scheduling step of the application. Schedulers specially designed for hybrid clouds simply ignore the inaccuracies of the given estimates and end up producing non-robust, low-performance schedules, which makes them sensitive to the uncertainties stemming from using these networking tools. This thesis introduces a proactive procedure to provide a certain level of robustness for schedules derived from schedulers that were not designed to be robust in the face of uncertainties of bandwidth estimates stemming from using unreliable networking tools. To make non-robust schedulers into robust schedulers, the procedure applies a deflation on imprecise bandwidth estimates before being used as input to non-robust schedulers. By proposing the use of refined (deflated) estimates of the available bandwidth, non-robust schedulers initially sensitive to these uncertainties started to produce robust schedules that are insensitive to these inaccuracies. The effectiveness and efficiency of the procedure in providing robustness to non-robust schedulers are evaluated through simulation. Schedules generated by induced-robustness schedulers through the use of the procedure is compared to that of produced by sensitive schedulers. In addition, this thesis also introduces a flexible scheduler for a special case of scientific applications modelled as a set of workflows grouped into ensembles. Although the novelty of this scheduler is the replacement of objective functions according to the user's needs, it is still a non-robust scheduler. However, the procedure was able to provide the necessary robustness for this flexible scheduler be able to produce robust schedules under uncertain bandwidth estimates. It is shown in this thesis that the proposed procedure enhanced the robustness of workflow schedulers designed especially for hybrid clouds as they started to produce robust schedules in the presence of uncertainties stemming from using networking tools. The proposed procedure is an important tool to furnish robustness to non-robust schedulers that are originally designed to work in a computational environment where bandwidth estimates are very likely to vary and cannot be estimated precisely in advance, bringing, therefore, improvements to the executions of scientific applications in hybrid clouds (AU) | |
| FAPESP's process: | 12/02778-6 - Dynamic Scheduling of Multiple Workflows for SaaS/PaaS Cloud Providers Considering Two SLA Levels |
| Grantee: | Thiago Augusto Lopes Genez |
| Support Opportunities: | Scholarships in Brazil - Doctorate |