Publicação
Power aware scheduler for heterogeneous environments
| Resumo: | Efficient or green computing is becoming a key issue in current programming techniques, going beyond high performance computing, by simultaneously considering issues such as energy or power consumption. In heterogeneous environments, where different processors and accelerators co-processors may coexist, there is a real opportunity to reduce the overall energy consumption of the system by using scheduling decisions in run-time that can have a good and quick response to changes in the different components. Current tools to aid the development of efficient applications lack yet these run-time facilities. This motivated the development of a new framework with a power-aware scheduler for heterogeneous environments, PASH-Frame, whose prototype is the key object of this dissertation. This work extended previous performance-based scheduling work to include run-time power-aware features, adding tools to measure power consumption at each device and using different scheduling decisions to get the best outcome according to pre-defined targets by the end user. To evaluate the overall behaviour of PASH-Frame, several tests were performed: 1000 SAXPY tasks with vector sizes varying from 16 thousand elements to 256 thousand; 200 SGEMM tasks with matrices varying from 64 thousand elements to 16 million and finally a test that combines the two previous ones. Results show that the scheduling algorithm implemented in the framework can achieve good results in some cases, in spite of not being able to make some critical decisions when it comes to energy consumption reduction like forcing a component to idle to save energy. |
|---|---|
| Autores principais: | Magalhães, João Paulo Fontoura Moutinho |
| Ano: | 2017 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Efficient or green computing is becoming a key issue in current programming techniques, going beyond high performance computing, by simultaneously considering issues such as energy or power consumption. In heterogeneous environments, where different processors and accelerators co-processors may coexist, there is a real opportunity to reduce the overall energy consumption of the system by using scheduling decisions in run-time that can have a good and quick response to changes in the different components. Current tools to aid the development of efficient applications lack yet these run-time facilities. This motivated the development of a new framework with a power-aware scheduler for heterogeneous environments, PASH-Frame, whose prototype is the key object of this dissertation. This work extended previous performance-based scheduling work to include run-time power-aware features, adding tools to measure power consumption at each device and using different scheduling decisions to get the best outcome according to pre-defined targets by the end user. To evaluate the overall behaviour of PASH-Frame, several tests were performed: 1000 SAXPY tasks with vector sizes varying from 16 thousand elements to 256 thousand; 200 SGEMM tasks with matrices varying from 64 thousand elements to 16 million and finally a test that combines the two previous ones. Results show that the scheduling algorithm implemented in the framework can achieve good results in some cases, in spite of not being able to make some critical decisions when it comes to energy consumption reduction like forcing a component to idle to save energy. |
|---|