Multi-Core Software

  Section 6 of 9  

Accelerating Video Feature Extractions in CBVIR on Multi-Core Systems

CONCLUSION

CBVIR is becoming one of the best solutions to retrieve useful information from today's massive amount of video data. To accelerate CBVIR on multi-core systems, we optimize and parallelize a set of representative visual feature extraction workloads in CBVIR. We analyze their scalability and memory performance on an 8-core system and draw several conclusions.

Firstly, we choose appropriate parallel schemes for the applications in CBVIR. Exploring different levels of parallelism and choosing the most favorable are necessary to enable optimal performance on multi-core systems. Secondly, we incrementally optimize the parallel performance by mitigating the parallel performance limiting factors, e.g., load imbalance removal, designing cache-friendly data structures, using different thread-scheduling policies, etc. Thirdly, we find most of the CBVIR applications have very good scaling performance. The main scalability limiting factors for SIFT and Gabor are load imbalance and the amount of available system bandwidth. Finally, the CBVIR system is significantly accelerated on multi-core systems and offers enhanced performance to satisfy user requirements.

  Section 6 of 9  

Back to Top