Acesso paralelo a arquivos em sistemas multiprocessadores. / Parallel access to files for multiprocessor systems.

AUTOR(ES)
DATA DE PUBLICAÇÃO

2010

RESUMO

This work presents a study on disk access bottlenecks for multiprocessor architecture systems and also proposes a specification to solve this problem and its implementation. The developed strategy is called FPA. The developed strategy makes use of shared memory architecture and uses shared data buffers among several application threads and also tries to optimize disk access time. In order to decrease disk concurrency, is allowed to just one application thread to read/write simultaneously and processing thread operations run parallelly to disk operations. Write operations are performed by an internal thread, which is started and managed by FPA. Several tests were performed concerning different amounts of application threads instances, file sizes and computing instructions. Results show performance gain with FPA in several situations. The main factor that affects performance is the amount of computing instructions on each application. In general, read operations showed a better performance compared to use of streams in situations where there is few computing instructions in the application, and write operations showed advantages on situations where there are more computing instructions in the application. Write operations are advantageous using data writing and computing instructions overlap. However, in cases of a large amount of computing instructions compared to disk operations, the time to perform disk operations becomes too few compared to computing time, occurring FPA performance loss compared to streams.

ASSUNTO(S)

gargalo no acesso a disco disk access bottleneck memory mapped files mapeamento de arquivos em memória

ACESSO AO ARTIGO

Documentos Relacionados