310 likes | 477 Views
Analysis of Multimedia Workloads with Implications for Internet Streaming. Lei Guo 1 , Songqing Chen 2 , Zhen Xiao 3 , and Xiaodong Zhang 1 Presented by: Zhen Xiao 1 College of William and Mary 2 George Mason University 3 AT&T Labs – Research
E N D
Analysis of Multimedia Workloads with Implications for Internet Streaming Lei Guo1, Songqing Chen2, Zhen Xiao3, and Xiaodong Zhang1 Presented by: Zhen Xiao 1College of William and Mary 2George Mason University 3AT&T Labs – Research The 14th International World Wide Web Conference
HTTP file Multimedia: Downloading Web Browser Web Server Long start-up latency Potential waste of traffic Media Player
HTTP Multimedia: Pseudo Streaming Web Browser Web Server Media Player also called progressive downloading or progressive playback
HTTP RTSP/MMS/HTTP meta file RTP/RTCP Multimedia: Streaming Web Browser Web Server Media Player Streaming Server
Goals and Objectives • How is multimedia content delivery doing in practice? • Streaming has many advantages over downloading for multimedia traffic • But what percentage of multimedia traffic is delivered via streaming? • What are the implications of different content delivery methods for multimedia traffic? • bandwidth efficiency, playback quality, etc. • Can we quantify the actual benefits of a streaming service? • What can we do to improve the current content delivery practice for multimedia traffic?
Existing Work • Streaming/Web sites • A small number of popular servers • No study on large number of Web sites • Clients • Educational [USITS01][NOSSDAV01] or enterprise environments [NOSSDAV02] • Very few study on commercial workloads • Data Sources • Pre-stored video objects [MMCN98], server logs [NOSSDAV02] • No flow level information • Focuses • Object popularity and sharing patterns, client interactivity [WWW01][ICDCS05] • Few on content delivery methods
Our Contributions • Analyze two large commercial multimedia workloads • Much larger scale • More detailed information (e.g., byte counters) • Focus on multimedia delivery methods, bandwidth efficiency, playback quality • Design and simulation of the AutoStream system • Provide streaming service for standard Web servers • Share the cost of streaming service among content providers
Outline • Background • Trace Collection and Processing • Workload Analysis • AutoStream • Conclusions and future work
Trace Collection • Two packet level media workloads collected with the Gigascope appliance • Server Workload: a large number of commercial Web sties hosted by a major ISP (a Web server farm) • Client Workload: a large group of home users connected to the Internet via a well-known cable company (cable clients) • The two workloads are independent! • 24 hour duration: 06/15/2004 8pm – 06/16/2004 8pm • We collected: • The first IP packets of all HTTP requests and responses • The first IP packets of all RTSP and MMS control messages • Byte counters: the number of bytes transferred through each TCP/UDP connection per second • All HTTP based P2P traffic were carefully filtered out
Traffic Overview • Total data size: 100GB in gzip format • Server workload • 1,095,984 media requests/response pairs • 4,498 unique server IPs, 79,309 unique client IPs • Client workload • 579,693 media requests/response pairs • 13,110 unique server IPs, 7,906 unique client IPs
Trace Processing • Downloading • User-Agent in HTTP request is a Web browser • Content-Type in HTTP response is audio or video • application/multipart: based on 34 most popular suffixes for media files (e.g. .mp3, .mpeg, etc.) • Pseudo streaming • Subtle differences from downloading • User-Agent in HTTP request corresponds to a media player • Most streaming uses RTSP and MMS • HTTP based streaming is very small
Trace Processing (Cont’d) • Processing • Decoded most popular media formats: Windows, Real, and QuickTime • Extracted URL, media encoding rate, and playback time. • Requested traffic: Content-Length in HTTP response or media length and encoding rate extracted from RTSP/MMS messages • Transferred traffic: actually transferred data based on byte counters.
Outline • Background • Trace Collection and Processing • Workload Analysis • AutoStream • Conclusions and future work
Multimedia Delivery Methods Server Workload Client Workload
Object Size: Server Workload Media traffic is always dominated by large objects % Connections % Traffic
Object Size: Client Workload % Connections % Traffic
Early Terminated Connections % Connections % Traffic Compared with downloading, clients using pseudo streaming tend to abort more and early, and hence cause less traffic.
20% Client access duration in streaming 11% 35% 44% Server Workload Client Workload Compared with downloading and pseudo streaming, clients using streaming are much more likely to terminate their access to an object earlier.
Bandwidth Efficiency Definition: The percentage of requested traffic that was actually transferred.
Rate Mismatch in Pseudo Streaming • Downloading rate: averaging the transferred bytes over the data transmission time. • Streaming rate: average object encoding rate Server Workload Client Workload Rate mismatch in pseudo streaming is common, which can cause the client to experience frequent delays in order to refill its buffer.
Advantages of Streaming • Rate adaptation • Transcoding • Multiple-Bit-Rate Encoding • Intelligent Streaming from Microsoft • SureStream from Real Networks • Frame thinning • Prioritized retransmissions • UDP based streaming • Server workload: RTSP: 10.4%, MMS: 23.5% • Client workload: RTSP: 26.8%, MMS: 21.5% • Only re-send lost packets that can arrive in time for the playback • Support for interactive operations • Pause, fast forward, rewind, etc.
Summary of Findings We found • Streaming is the most efficient approach for multimedia delivery. but • Most multimedia traffic is delivered via downloading. Why is streaming not widely used? • Streaming has associated expenses • Benefits not obvious AutoStream • Share the cost of streaming among content providers • Try streaming before you buy
Outline • Background • Trace Collection and Processing • Workload Analysis • AutoStream • Conclusions and future work
HTTP RTSP/MMS/HTTP HTTP RTP/RTCP HTTP AutoStream Overview server AutoStream Client server Try before you buy! Server Farm Existing Web servers can provide real streaming service!
AutoStream Architecture Client Request Handler Virtual Streaming Engine Windows Media Streaming Engine Real Media Streaming Engine QuickTime Media Streaming Engine Streaming Media Converter Prefix Cache Engine
Evaluation • Trace driven simulation using the server workload • On-demand cache with initial segments only. Cache cleaned hourly • When converting non-streaming traffic into streaming, assume the same access patterns as existing streaming traffic • When the available bandwidth to a client is lower than the encoding rate of the media, transcode the media into an appropriate lower encoding rate • Metrics • Traffic reduction due to changes in access patterns • Benefit of prefix caching is not included • Start-up latency reduction for downloading traffic
Traffic Reduction AutoStream reduces downloading traffic by 78% and pseudo streaming traffic by 72%. Without transcoding, the reductions are 56% and 43%.
Start-up Latency Reduction 63% sessions previously experiencing start up delays no longer do after AutoStream
Outline • Background • Trace Collection and Processing • Workload Analysis • AutoStream • Conclusions and future work
Conclusions and Future Work • We found that streaming has many benefits for delivering multimedia traffic, but only limited deployment. • We proposed AutoStream, a system to bridge the gap between the potential and practice in multimedia delivery. • Lesson learned – always do reality check • Don’t assume “If a technology is good, it’ll be used.” • Future work • Multimedia delivery in peer-to-peer systems • Streaming delivery quality on the Internet
Thank you! Questions?