Abstract
Data locality in distributed memories has a significant performance impact on NUMA multi-core systems owing to non-uniform memory accesses. In addition, memory contention also influences the performance of multi-core systems. The performance degradation caused by both effects should be analyzed before performance optimization because data locality and memory contention are mutually dependent. A reduction of one effect could cancel performance gains due to another effect. A novel post-processing method based on non-intrusive tracing is proposed in this paper to analyze the performance impact incurred by both data locality and memory contention in a quantitative, comparable way. It makes use of non-intrusive tracing, which has no impact on normal execution and timing. The analysis provides results including data locality and memory contention penalties, which can be used as a reference to improve performance.
Item Type: | Conference or Workshop Item (Report) |
---|---|
Faculties: | Mathematics, Computer Science and Statistics > Computer Science |
Subjects: | 000 Computer science, information and general works > 004 Data processing computer science |
Language: | English |
Item ID: | 47380 |
Date Deposited: | 27. Apr 2018, 08:12 |
Last Modified: | 13. Aug 2024, 12:54 |