1. Differentiated Multimedia Web Services Using Quality Aware Transcoding S. Chandra, C.Schlatter Ellis and A.Vahdat InfoCom 2000, IEEE Journal on Selected areas in Communications, Dec 2000

2. Outline Introduction Introduction Quality Aware Transcoding Quality Aware Transcoding Experiment objective and design Experiment objective and design Results Results Conclusion Conclusion

3. Introduction Web service to provide low-latency access to its contents is constrained by available network bandwidth Web service to provide low-latency access to its contents is constrained by available network bandwidth Introduce a mechanism for managing network resources at the application level Introduce a mechanism for managing network resources at the application level Use transcoding to allow web servers to customize the size of objects on a webpage by dynamically varying the size of multimedia objects on a per-client basis. Use transcoding to allow web servers to customize the size of objects on a webpage by dynamically varying the size of multimedia objects on a per-client basis. Quality vs. Size tradeoffs of JPEG images Quality vs. Size tradeoffs of JPEG images Evaluate performance of several bandwidth management policies Evaluate performance of several bandwidth management policies

4. Introduction It is possible to use informed transcoding techniques to provide differentiated service and to dynamically allocate bandwidth among different client classes, while delivering a high degree information content for all clients. It is possible to use informed transcoding techniques to provide differentiated service and to dynamically allocate bandwidth among different client classes, while delivering a high degree information content for all clients.

5. Quality Aware Transcoding Focus on transcode reduce bandwidth requirements on the server Focus on transcode reduce bandwidth requirements on the server Quantify the quality vs. size tradeoff for transcoding JPEG images Quantify the quality vs. size tradeoff for transcoding JPEG images Changing JPEG compression metric: JPEG Quality Factor, correspond to quality loss Changing JPEG compression metric: JPEG Quality Factor, correspond to quality loss Use adjusting JPEG compression metric as the informed transcoding technique, this result is equally valid for any transcoding with well understood tradeoff characteristics Use adjusting JPEG compression metric as the informed transcoding technique, this result is equally valid for any transcoding with well understood tradeoff characteristics

6. Experiment objective and design Objective: Objective: bandwidth consumption control bandwidth consumption control differentiated service for preferred and ordinary clients differentiated service for preferred and ordinary clients System Architecture System Architecture

7. Experiment objective and design Performance measures: Performance measures: Bandwidth Consumed maintain within pre-defined level Bandwidth Consumed maintain within pre-defined level Image Quality Factor tradeoff quality for size retain as much information quality as possible Image Quality Factor tradeoff quality for size retain as much information quality as possible Client Access latency acceptable client latency, 10s as threshold Client Access latency acceptable client latency, 10s as threshold

8. Experiment objective and design Web Service Policies Web Service Policies Traditional without bandwidth control mechanisms. Use the modified Apache Web server Traditional without bandwidth control mechanisms. Use the modified Apache Web server Modbandwidth bandwidth limiting web service that prioritizes the network packets, delaying packets with low priority requests (Apache mod_bandwidth module) Modbandwidth bandwidth limiting web service that prioritizes the network packets, delaying packets with low priority requests (Apache mod_bandwidth module) Denial temporarily denies requests under heavily loaded conditions (HTTP error code 503 : Service Unavailable) Denial temporarily denies requests under heavily loaded conditions (HTTP error code 503 : Service Unavailable)

9. Experiment objective and design Transcoding transcode an image to a number of variations Ensure that any loss in image quality is efficient, definded as a transcoding that loses at least as much in image size as the loss in image information quality Transcoding transcode an image to a number of variations Ensure that any loss in image quality is efficient, definded as a transcoding that loses at least as much in image size as the loss in image information quality

10. grayscale 85 KB 10 KBOrigin 116KB Crop 40KB

11. 5 % quality 16% size Original 17KB 50% quality 49% Size Effective Ineffective

12. Experiment objective and design Experimental Workload Experimental Workload Access Traces Access Traces client access trace that captures request arrival times. client access trace that captures request arrival times. Analyzing access made via the NLANR proxy caches, use the accesses to geocities.com Analyzing access made via the NLANR proxy caches, use the accesses to geocities.com Exp trace: changes access rate every 30 seconds, scale trace to 12500 seconds(3.5hrs) Exp trace: changes access rate every 30 seconds, scale trace to 12500 seconds(3.5hrs) Image Collection Image Collection 3531 JPEG images total 38 MB 3531 JPEG images total 38 MB

17. Experiment objective and design Client Classes Client Classes modify http_load to generate a configurable percentage of requests from different client classes modify http_load to generate a configurable percentage of requests from different client classes http_load informs the server about the class that a request belongs to using custom HTTP headers http_load informs the server about the class that a request belongs to using custom HTTP headers Web Server Web Server Apache server computes the bandwidth trend by averaging over the past 30 mins interval Apache server computes the bandwidth trend by averaging over the past 30 mins interval

18. Experiment objective and design Server implements Transcoding mechanism: If the average consumed bandwidth for the past half hour is more than target bandwidth, the server proportionately lower quality variations of images. When the consumed bandwidth exceeds twice the target bandwidth, the server denies further requests. Server implements Transcoding mechanism: If the average consumed bandwidth for the past half hour is more than target bandwidth, the server proportionately lower quality variations of images. When the consumed bandwidth exceeds twice the target bandwidth, the server denies further requests. Client: http_load to simulate access compute the individual access latencies play back the client access trace generated from the NLANR proxy traces. Client: http_load to simulate access compute the individual access latencies play back the client access trace generated from the NLANR proxy traces.

19. Results Bandwidth Control Bandwidth Control Target bandwidth: 200KB/s Target bandwidth: 200KB/s Avg. access rate: 5.7 accesses/s Avg. access rate: 5.7 accesses/s

20. Overshoot the target bandwidth : Traditional: 65% of time Modbandwidth: 0% Denial: 35% Transcoding: 45% Traditional and Denial provides best control over bandwidth, but driving client latencies to unacceptable level

22. Server latencies Denial requests : infinite latency and Quality factor of 0 Transcoding scheme gracefully degrades the Quality factor

23. Modbandwidth: 90% of requests takes 10s to service Denial, Transcoding: less than 1s

24. Transcoding provides useful bandwidth control by gracefully degrading image quality factors. Transcoding provides useful bandwidth control by gracefully degrading image quality factors. Transcoding does not control bandwidth effectively in the presence of flash crowds and if there ’ s a wide differential between target and requested bandwidth. Transcoding does not control bandwidth effectively in the presence of flash crowds and if there ’ s a wide differential between target and requested bandwidth. Denial and Modbandwidth control their bandwidth consumption by denying service for a large portion of the images Denial and Modbandwidth control their bandwidth consumption by denying service for a large portion of the images

25. Results Differentiated Web Service Differentiated Web Service Since Modbandwidth performs with unacceptable latencies, only compare Denial and Transcoding schemes Since Modbandwidth performs with unacceptable latencies, only compare Denial and Transcoding schemes Avg. access rate: 28.6 accesses/s Avg. access rate: 28.6 accesses/s Target bandwidth : 1MB/s Target bandwidth : 1MB/s Avg. bandwidth demanded: 1.5MB/s Avg. bandwidth demanded: 1.5MB/s 40% of available bandwidth for the preferred clients 40% of available bandwidth for the preferred clients 20% of the clients be preferred clients 20% of the clients be preferred clients

26. Prefer clients: Denial and Transcoding: 90%,65% less than 400KB/s

27. Target bandwidth is closer to Max. bandwidth of Transcoding, the system better handles the heavy load at time =4500s

29. Conclusion Explore a web service that uses informed transcoding to manage its bandwidth consumption Explore a web service that uses informed transcoding to manage its bandwidth consumption The service degrades the quality gracefully for different user classes, while still managing overall consumed bandwidth effectively. The service degrades the quality gracefully for different user classes, while still managing overall consumed bandwidth effectively. Allow server to manage its bandwidth without adding excessive latency or denying service Allow server to manage its bandwidth without adding excessive latency or denying service Future work: to specify the relative importance of various multimedia components of web pages. Future work: to specify the relative importance of various multimedia components of web pages.