Wednesday, May 14, 2008

On dear peer-to-peer

I will be presenting a paper* at the IEEE IwQoS early next month containing the analysis of a large scale peer-to-peer live video multicast streaming session on the Internet. Think of the P2P video multicast system as the Bittorrent for video streaming (instead of file-sharing). The system was sending a video stream of a baseball match to 10s of thousands of viewers on the Internet using P2P technology.

The presentation and paper are available online. Here are two results from the paper that in my opinion warrant particular notice.

Figure 1 shows the aggregate download and upload bandwidth consumed by the P2P system. Note the scale on the Y axis - Gbps! I am wondering, this is only 1 video stream. What happens to the Internet when 1000s of such streams become available online? Were networks designed for such usage?

An interesting artifact in this figure is that the aggregate download rate of all peers exceeds the aggregate upload rate of all peers. The difference was made up through "bandwidth injecting super-servers" of the content provider. Still, it is absolutely remarkable that the amount of additional bandwidth required is almost constant even as the number of peers increase (Figure 2).
Figure 1: Total bandwidth. Click to enlarge.

In Figure 2, you can see the number of concurrent peers in the P2P system over the time-period of the streamed game (hour 4 to hour 8). Look at the rate of change of peers in (peers joining, peers leaving) the P2P system. Keeping in mind that most of the bandwidth comes from these very peers, it is remarkable that this highly dynamic pool of peers is able to sustain the P2P system. Things get very exciting at the end of the baseball game (Hour 8): Everybody wants to leave. Now that is a big challenge for any P2P system.
Figure 2: Peer Dynamics. Click to Enlarge.

*Joint work with Jatinder Pal Singh (T-Labs) and Aditya Mavlankar, Pierpaolo Baccichet, and Bernd Girod (Stanford University)

No comments: