Defesa de Dissertação de Mestrado


19/01/2022 - 12h11

DEFESA DE DISSERTAÇÃO DE MESTRADO – Programa de Pós-Graduação em Ciência da Computação


ALUNO: Angelo Elias Dalzotto

ORIENTADOR: Dr. Fernando Gehm Moraes

COORIENTADOR: Dr. Marcelo Ruaro

BANCA EXAMINADORA: Dr. Everton Alceu Carara (PPGCC/UFSM), Dr. Ney Laert Vilar Calazans (PPGCC/PUCRS)

DATA: 24 de fevereiro de 2022

LOCAL: Videoconferência

HORÁRIO: 10:00

Link para acessar videoconferência

Senha: 1234

The increasing core count in many-core systems introduced management challenges, including scalability, portability, and reduced overhead to user applications. Works available in the literature have their management tightly coupled to the many-core operating system. The coupling implies low modification flexibility of the management organizations and less portability. The state-of-the-art also shows that few works proposed different management organizations, only exploiting an organization to evaluate the quality of a single objective, such as power or temperature. The present work proposes a management framework called Management Application (MA), loosely coupled to the platform. The MA transforms the management into a distributed application, benefiting from the parallel processing power of many-cores. Compared to cluster management, the costs and benefits to manage a benchmark with real-time constraints using the MA revealed improved memory footprint and higher management throughput due to parallelization imposed by the MA. It is also proposed a mapping heuristic that uses virtual clusters to reduce the execution cost keeping a centralized view of the system, and with a built-in defragmentation procedure. Results are evaluated against a state-of-the-art heuristic in clustered and per-application management, revealing reduced distance between communicating tasks and similar heuristic execution time to the clustered approach. The defragmentation can decrease the distance between communicating tasks with few migration movements. The framework structure is optimized with a broadcast-based network, used for exchanging small management messages, and a monitoring structure that exploits the broadcast-based network and the direct memory access to reduce the monitoring cost. The network reduces the interference in user applications and the execution time, while the monitoring structure allows smaller management latency. Lastly, the MA paradigm is generalized to a RISC-V processor, reducing the number of executed instructions and the memory footprint. The final result is a many-core platform that implements the MA organization with a state-of-the-art processor.