As compared to a complex single processor based system, on-chip multiprocessors are less complex, more power efficient, and easier to test and validate. In this work, we focus on an on-chip multiprocessor where each processor has a local memory (or cache). We demonstrate that, in such an architecture, allowing each processor to do off-chip memory requests on behalf of other processors can improve overall performance over a straightforward strategy, where each processor performs off-chip requests independently. Our experimental results obtained using six benchmark codes indicate large execution cycle savings over a wide range of architectural configurations.
|Original language||English (US)|
|Number of pages||8|
|Journal||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|State||Published - 2004|
ASJC Scopus subject areas
- Theoretical Computer Science
- Computer Science(all)