1. Differentiated Multimedia Web Services Using Quality Aware Transcoding Surendar Chandra, Carla Schlatter Ellis and Amin Vahdat Department of Computer Science, Duke University

2. Jinho Chang in VC Lab. Abstract A robust mechanism for managing network resources at the application level Use informed transcoding techniques to provide differentiated service and to dynamically allocate available bandwidth among different client classes

3. Jinho Chang in VC Lab. Introduction A primary goal of a web service is to provide low-latency access to its contents is constrained by available network bandwidth. Caching at web proxies is one traditional technique, but much web content is dynamically generated (maps, stock charts) or un-cacheable(sites selling access to images or movies).

4. Jinho Chang in VC Lab. Trends that illustrate the need for a differentiated QoS system: E-commerce maintain better QoS for preferred customers Web Hosting Maintain web pages on behalf of their customers. Charge their customers based on the size of the web site and the aggregate consumed bandwidth

5. Jinho Chang in VC Lab. Differentiated service enables: Web services to dynamically allocate the available bandwidth among different user classes Subscrioption services to provide different versions of contents to clients based on customer status. Web hosting services to share their bandwidth for different classes of hosted clients E-commerce sites to allocate their bandwidth to customers who are making a sell Flexibility to redirect unused preferred resources to non-prefered customers About 77% of the data bytes accessed are from multimedia objects. Of these, 67% of the data are transferred for images.

6. Jinho Chang in VC Lab. Use transcoding to dynamically create variations of same multimedia object (multimedia objects are amenable to soft access through a quality-versus- size tradeoff) Inherent tradeoff characteristics: Information quality loss Computational overhead Potential benefits of reduced bandwidth requeirements

7. Jinho Chang in VC Lab. BackGround: Quality Aware Transcoding Transcoding operations are ofter performed to fit an object ot the characteristics of the display device. Images have been transcoded to thumb-nails, gray-scale, progressive formats as well as transcoded to textual information Ex: Full color JPEG images are transcoded to bitmap form to reduce latency for modem user.

8. Jinho Chang in VC Lab. Information quality loss directly corresponds to the change in th JPEG Quality Factor the JPEG Quality factor change based on the sum of Minimum Code Unit(MCU) block counts for all the different color space components Images with high coefficients for low frequency components as well as images with initial JPEG Quality factor greater than 80 can transcode images efficiently at a better percentage than the base case.

9. Jinho Chang in VC Lab. The transcoding required to provide differentiated service for static images is performed once and the results of the transcodeing are cached for reuse. Concerned with the limited, expensive network bandwidth to the Web server, not the transcoding computational and storage overhead. Experiment Design

10. Jinho Chang in VC Lab.

11. Performance Measures Bandwidth consumed Image Quality Factor Client Access Latency Web Services Policies (Apache Web server) Traditional Mod_bandwidth Denial Transcoding Experimental Workload Access traces:from NLANR Image collection: 3531 JPEG images,total 38MB Client classes:Http_load

12. Jinho Chang in VC Lab. Mesured in 5 minute intervals, an average traffic of 0.2 acesses/sec High accesses during evening hours and lower accesses during early morning hours

13. Jinho Chang in VC Lab. Images are of high quality. 60% of the images have JPEG Quality factor higher than 70

14. Jinho Chang in VC Lab. 40% of images are bigger than 10KB

15. Jinho Chang in VC Lab. Transcode efficiently when images are transcoded to a Quality factor of 30% or less. Cache takes 165MB Transcode efficienctly(%)

16. Jinho Chang in VC Lab. Efficent: a transcoding that loses at least as much in image size as the loss in the image information quality

17. Jinho Chang in VC Lab. Computes the bandwidth trend by averaging over the past 30 minute interval 比 target bandwidth 大, 降低圖的品質. 超過 target bandwidth 兩倍,deny requests

18. Jinho Chang in VC Lab.

19. Results

20. Jinho Chang in VC Lab.

22. Differentiated Web Service Compare Denial and Transcoding schemes.

23. Jinho Chang in VC Lab. Target bandwidth 1MB,average bandwidth demended at 1.5MB

24. Jinho Chang in VC Lab.

25. Conclusion Explore a web service that uses informed transcoding to manage its bandwidth consumption Transcoding can allow the server to manage tis bandwidth without adding excessive latency or denying service Transcoding also allows web server to provide differentiated service by allocating its bandwidth for different usage classes.

26. Jinho Chang in VC Lab. Future work Currently investigating techniques allow web designers to specify the relative importance of various multimedia components of web pages.