Performance and Quality-of-Service Analysis of a Live P2P Video Multicast Session on the Internet

Transcription

1 Performance and Quality-of-Service Analysis of a Live P2P Video Multicast Session on the Internet Sachin Agarwal 1, Jatinder Pal Singh 1, Aditya Mavlankar 2, Pierpaolo Bacchichet 2, and Bernd Girod 2 1 Deutsche Telekom A.G., Laboratories & TU Berlin Ernst-Reuter-Platz Berlin, Germany 2 Department of Electrical Engineering, Stanford University 350 Serra Mall Stanford, CA 94305, USA June 2, 2008 Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

2 Outline Introduction to P2P live video streaming Background information about the log data Results via analysis of the log data Summary and Conclusions Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

3 Outline Introduction to P2P live video streaming Background information about the log data Results via analysis of the log data Summary and Conclusions Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

4 Introduction to P2P live video streaming Uses P2P technology to stream video content Several deployed commercial implementations today Increasing content/channels becoming available Bulk of the bandwidth comes from broadband connections: a low-cost content delivery technology (?) Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

5 Outline Introduction to P2P live video streaming Background information about the log data Results via analysis of the log data Summary and Conclusions Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

6 The analyzed P2P video streaming system Live multicast of a 4-hour baseball game to over 120K IP addresses on the Internet Pre-planned live multicast, therefore additional bandwidth provisioning was possible Content format: CBR, WMV codec, 759 kbps audio+video stream at 29fps and VGA resolution Content generated and mostly consumed in developed country, hence, good network infrastructure Mesh-based P2P algorithms used Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

7 P2P Client P2P client was built into Internet Explorer (ActiveX plug-in) WMP as decoder P2P clients freely distributed (no access control, no DRM) The local nature of content resulted in mostly local installations... But 51 countries had one or more clients Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

8 Logging P2P clients logged information to a centralized logging server Information included client-id, received time-stamp, IP address, buffer state, bytes sent, bytes received, running time, firewall type, channel information, recently lost packets, etc. Clients logged information every 5 minutes, or when there was a special event (start, stop, etc.) Challenge was in rebuilding the macro-characteristics of the multicast session from these log snippets Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

9 Outline Introduction to P2P live video streaming Background information about the log data Results via analysis of the log data Summary and Conclusions Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

10 P2P system size and peer dynamics Peers in system x Time (hours) Joins/leaves per second second moving average Maximum join & leave rate = 80 & 328 peers per second respectively 40 Joins 20 Leaves Time (hours) Figure: Peer churn: The total number of peers in the system vs. time. Non-linear and highly dynamic system Flash crowd-like effects Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

11 Peer persistence and churn Table: Statistics of peer persistence in the system (seconds). Minimum 1 1st Quartile 18 Median 106 Mean rd Quartile 1205 Maximum Why the short peer persistence? Short attention span, bad quality? Note: More than connecting peers did not receive any stream blocks, probably triggering an early exit Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

12 Huge bandwidth requirement at the network edge This is not a pure P2P system, instead, P2P+CDN Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27 P2P and CDN bandwidth Download Upload Difference 30 Rate (Gbps) Time (hours) Figure: Network bandwidth load: The aggregate download and upload data-rates summed over all peers.

13 Fewer net bandwidth contributors Most uplink connections less than 759 kbps Incentives to encourage more bandwidth contribution? Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27 Peer bandwidth: contribution vs. usage Net contributing peers Net receiving peers Number of peers log 10 (Ratio of downloaded to uploaded bytes) Figure: Histogram of the logarithm of the ratio of bytes downloaded to bytes uploaded by peer clients.

14 Type of Internet connections Table: Statistics of the type user premises termination technology. Most peers were behind a NAT. Direct connections 8.98 UPNP 6.53 Firewall 6.80 NAT Other/Unknown 7.44 Effective P2P streaming systems need effective NAT traversal capabilities. Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

15 IP address subnet analysis Table: IPv4 address distribution in terms of IP-subnet mask prefixes. Unique Prefix 8 bits 16 bits 24 bits Total Statistics of IP addresses per prefix Minimum st Quartile Median Mean rd Quartile Maximum Highly asymmetric distribution of the 120K IP addresses across different subnet sizes Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

16 IP address ISP classification Table: Classification of the 731 participating Autonomous Systems (ASes) AS type Number found T-2 (Small ISPs) 314 COMP (Customers) 90 NIC (National Information Centers) 87 EDU (Universities) 55 T-1 (Large ISPs) 20 IX (Internet Exchange Points) 1 Not Classified 198 P2P system interacts with hundreds of different ISPs Classification based on [DIMITRO] 1 1 X. Dimitropoulos, D. Krioukov, G. Riley, and KC Claffy, Revealing the autonomous system taxonomy: The machine learning approach, In Proc. of Passive and Active Measurements (PAM), Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

17 Sparse peer population in many ASes results in significant inter-as traffic Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27 Number of IP addresses per AS 39% < 1% 6% 22% % Figure: Peer IP addresses per AS. Although just two ASes accounted for more than half of all peer IP addresses, many other ASes contained very few IP addresses.

18 Delivered quality (stream blocks) Mean = , Median = Frequency Ratio of blocks received by WMP to blocks needed by WMP Figure: Histogram of ratio of blocks received by the media decoder to blocks needed for perfect playback for the first 5 minutes on peers. Significant variance in delivered quality Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

19 P2P live video streaming Table: Statistics of the video quality ratio for the first 5 minutes at the peers (1 is best). Minimum st Quartile 0.88 Median 0.95 Mean rd Quartile 0.98 Maximum 1 Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

20 Burst losses Average number of consecutive lost blocks: Min = 0, Max = >40,000 peers do not lose any blocks Frequency Average number of consecutive lost blocks Figure: Histogram of the average number of consecutive blocks never delivered to peers during the first five minutes. Multiple consecutive blocks are lost at times Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

21 Quality variation across ASes - lost blocks WORSE QoS BETTER QoS Number of ASes Average fraction of correctly received stream Figure: Histogram of the average fraction of correctly delivered stream blocks to ASes with 10 or more peers. Different ASes receive very different QoS in terms of correctly delivered stream blocks Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

22 Channel startup time Number of peers Startup delay [sec] Figure: Histogram of the channel startup time Slow slow slow (The average was 32 seconds) Live P2P streaming cannot take advantage of tricks like preemptive caching Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

23 Quality variation across ASes - channel startup time BETTER QoS WORSE QoS Number of ASes Average start up time Figure: Histogram of the average channel startup time on peers in ASes. Different ASes receive very different QoS in terms of channel startup times Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

24 Outline Introduction to P2P live video streaming Background information about the log data Results via analysis of the log data Summary and Conclusions Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

25 Summary P2P video multicast is a bandwidth-intensive application Peer behavior was highly dynamic The total bandwidth contribution of peers fell short of the total required bandwidth Most peers connected from behind Network Address Translation (NAT) devices The peers were widely distributed: 51 countries and 731 Internet autonomous systems (ASes) Significant variance in the fraction of correctly delivered stream (and hence, delivered quality) and channel startup time across different Internet ASes Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

26 Conclusions Access networks may get congested due to P2P video streaming: access networks are expensive to scale Greater demand from the network that P2P file-sharing; P2P file-sharing is less QoS intensive Designing traffic localization algorithms: not simple due to the number of ASes involved Significant scope for algorithmic improvements in stream delivery quality and channel startup times Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27

27 Q U E S T I O N S? Download Upload Difference 30 Rate (Gbps) Time (hours) Agarwal et al. (T-Labs, Stanford U) P2P multicast analysis June 2, / 27