Tuesday, May 22, 2007

P2P Streaming and Long Tail: Like Oil on Water

...or Misunderstanding the Long Tail

Here is another problem I have with P2P streaming. It purportedly delivers long-tail content.

I have already discussed how P2P streaming does not work. In fact, it does work, well… kinda, sorta, but it is “the weakest link” among all streaming and all P2P technologies. I believe that means it is destined to end up on the fossil heap thanks to the propensity of the marketplace toward natural selection of better technologies. The 40% extra bandwidth usage required for P2P streaming, and the fact that most connections are asymmetric will kill any chance that P2P streaming will actually work in the marketplace. No ISP will incur 40% additional costs on your behalf even if you are a brilliant billionaire (okay that was a low jab, but appropriate).

Let us for the moment assume that Joost manages to circumvent the laws of either Physics or Economics, and manages to make P2P streaming technology work. Let us assume that Joost or Vudu, or their users are able to convince ISPs to change their offering – make connections symmetric and allow P2P on their networks. (Unlikely – even the University of Ohio has banned P2P on their campus, despite a lot of criticism – but bear with me). Or that Vudu has some technology (Voodoo?) up their sleeve that will make P2P streaming consume a lot less bandwidth. Again, we would have to rewrite our college Physics books, but really have faith, will you? :-) To be fair, Vudu may use P2P streaming only for quick starting of streams, and not for the entire stream.

So, now that we have P2P streaming, let us put some Long Tail content out there to stream. Let us say someone (lets call him Mustafa from San Diego) connects to the Turkish Television Channel on Joost and brings up a Turkish movie. Now, in order to serve this customer in the P2P streaming model, we need peers! Someone, anyone, watching the same movie at about the same time (or within the 2 hours built-in to the buffer)? In an on-demand world? Not too many! Besides Akimbo has cornered the market on Turkish Television! (Sorry, another low jab).

OK, lets suspend our belief one more time. Let us say there are a couple of other people watching the exact Long Tail content within plus or minus one hour from Mr. Mustafa. Mr. Taimur (and of course has a “lame” 400kbps connection) is on the same ISP (defying all logic, he lives right next to Mr. Mustafa) but Mr. Ataturk is on a different ISP (it doesn’t matter where he lives). If you look at the diagram below (extremely simplified for all you “non-technical” people) you will see that Mr. Mustafa and Mr. Taimur connect through the same router or DSLAM, or whatever their ISP uses at the POP.

Unfortunately as you can see, Mr. Ataturk connects to this peer network through the Internet and saves no bandwidth (nor does his ISP). Of course, his ISP is extremely unhappy since all it means is more bandwidth usage. So if you were the ISP, what would you do? I suggest that you have two choices – cap the bandwidth or charge more for it.

In other words, the P2P streaming provider has increased the bandwidth (in comparison to straight streaming delivered from a server) and has then shifted the bandwidth off their network on to that of the users’ ISPs. If you ran an ISP, would you put up with it?

Actually the news is even worse. Most Telco based ISPs use DSLAMs, and the traffic shaping is usually done on the DSLAM itself. So even though they are on the same ISP, Mr. Mustafa and Mr. Taimur, probably would not be able to do much for each other in P2P streaming.

We have not even talked about the QOS (Quality of Service) issues of Streaming Video which of course apply to P2P Streaming Video. I have argued ad nauseum that Streaming quality cannot be guaranteed on the public Internet because of all the nodes in between over which no one has control (and Joost’s technical paper admits that). And in the above example, we have assumed that all the peers are fairly close – what if the majority of these peers lived across the Atlantic? QOS would be non-existent – you can’t wish away latency at those distances.

Of course, Joost could forget about the long tail and focus more on the Fat Tail (the other end of the graph where the big numbers are). Yes, that would work, wouldn’t it? It might even work with P2P streaming assuming some of the physics were taken care of. Wait! There is another technology that serves the Fat Tail quite well – its called Cable TV!

Incidentally, L3 Technology works with both the Long Tail and the Fat Tail. And it uses P2P but in a download and play scenario. I will write next about how P2P really works in a real word scenario. Stay Tuned!


lex said...

Nobody is talking about using pure P2P for internet VoD. The point is that P2P can be used to reduce server load for the very popular movies. Unpopular ("long-tail") movies will still hit the central server, but more of them will be serviceable with fixed resources due to the offload of the popular movies to the P2P sharing system.

If P2P is used only to help scale the central server, how can it be worse?

If UX is missing its targets, fail back to the central server.

As to the cross-network bandwidth issues; cross-network bandwidth can be reduced using locality mechanisms. For example, the network coordinates scheme currently deployed in Azureus. Place sharing peers into persistent groups with network locality.

Anil Gupte's Video Blog said...

Thanx Lex - you exactly made my point. For popular movies, we have a better solution for you than P2P streaming. It's called Cable TV - optionally with a PVR.

Your other points have also been addressed in the blog.

lex said...

I don't think I made that point. I claim that P2P is most suitable for the *popular* movies, to take load off the central servers, but is not suitable for the unpopular ones. The unpopular "long-tail" movies can be delivered through other, more traditional techniques, like cable TV.

Anil Gupte's Video Blog said...

Amazing! Deliver long-tail on Cable TV? I am speechless! :-)

lex said...

If the file access distribution follows a power law distribution, an auxiliary P2P system can take a lot of load off the centralized server components just by caching the first few files.