MLOOF: Arcabouço de descarregamento de processamento multi-nível

• Ano de defesa: 2020
• Autor(a) principal: Leandro Noman Ferreira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO; Programa de Pós-Graduação em Ciência da Computação; UFMG
• Palavras-chave em Português: Computação; Internet das coisas; Sistemas de comunicação móvel; Computação em nuvem; Computação – Teses; Internet das coisas – Teses; Sistemas de comunicação móvel – Teses; Computação em nuvem - Teses
• Link de acesso: http://hdl.handle.net/1843/36532

The term offloading indicates the action of changing the processing location of a computational activity. The purpose of using offloading is to reduce the processing time of applications, reduce the power consumption of the devices and eventually enable the execution of tasks that would not be possible on devices with reduced resources. This work presents an offloading framework called MLOOF (Multi-Level Online Offloading Framework). MLOOF has a three-level architecture composed of devices, cloudlets (intermediate servers) and cloud servers. The addition of cloudlets makes it possible to increase network throughput and reduce latency between device and server as they are closer to the devices. The code running on the device has a decision engine that chooses, when starting a method, where the method should be executed, whether on one of the servers or locally. The decision takes into account processing time prediction in the different execution environments and energy consumption of the device. The system was evaluated experimentally in a semi-controlled environment, using a testbed and dedicated servers in the cloud, and using a simulator. The results show that the three-level offloading framework strategy achieves the goals of reducing processing time and energy consumption, bringing great gains, especially when used by devices with less computational resources. In our experiments, the execution time using the cloudlets is up to 49% less than the execution on the cloud and the energy consumption when executing the code remotely remains constant, while when executing locally the consumption is exponential.