Slashdot has posted an article about a paper written by one of Napster's founding engineers. It is a mathematical evaluation of Gnutella discussing why the network won't be able to scale up to any reasonable size.

There's an interesting ongoing discussion at Slashdot. What do you think?

Slashdot Post/Discussion

Mathematical Evaluation of Gnutella

Tru dat!

Also, the essay is apparently over a year old, and I would assume that superpeer tech invalidates many of the assumptions therein.

yah... seriously misinformed...

And.. Napsters founding engineers for lack of a better word, weren't all that amazing. Their competitors at the time were Scour... comparatively Scour was EXPONENTIALLY better p2p technology even though they had far less time to get it off the ground.

It could scale to hundreds of thousands probably millions of simultaneus users ... Napster wasn't even close to that.

So anyway.. yah.

i'm not sure you're as informed as you think, spyed.

napster was the first of the bunch to top a million concurrent users (by far). the stats you saw in the client never reflected it though; they were only per-server.

Fletch:

As far as I know from what I was told by Napster executives, at the time they were shut down they were still working on connecting all of these simultaneus users.

So while they had millions of simultaneus users, they could not all search each other.. I think they could only search on their own server or perhaps one or two servers in the Napster server farm, but never all the servers at the same time.

Scour could do this from day one. Granted the code broke a dozen times and they had to rebuild it as the system would scale.. but it could always search all the users on it's network and by the time the polished SX product rolled out.. they could search all 250,000+ users in under one second.

No one did this before them; no one.

the only thing that was never linked on napster was chat. i think the order of what-was-linked-when was something like im -> hotlist -> browsing -> search, but all of those were linked before the shutdown.

regardless, this isn't a paper about searching across a network, this is a paper about scalability of the network's architecture. over the course of a couple years, napster's network grew from a single machine to over 1 million users quite smoothly (albeit also somewhat suprisingly =).

oh, and this is all pretty moot anyway because, as god node pointed out, this paper is just shy of a year old. i don't think there are very many gnutella networks around that aren't using some newer form of the model.

oh, this old article again. I thought we already found out Gnutella can't scale, but it doesn't care anyone since we can find everything in a typical horizon.... and Gnutella network as whole can grow grow grow.

so, Moak, you think is true?
it can't scale??

I believe maths don't fit the Internet everdeveloping reality, really

eyesbomb, why do we have TTLs? You can't reach more than you can reach in a so called 'horizon'. Everybods seems to be fine with that solution, you find all you need in a horizon. Superpeers might increase the horizon, but hoizons will stay. If scalable means everybody on earth can use Gnutella... yes they can.

I would second Moak on that one -- the goal is not necessarily to see all of the files of everyone on line at any one time. Rather, the goal is to see enough files to get what you want.

As an earlier poster mentioned, Gnutella is a far different network today than it was when this article was written, with changes such as LimeWire's UltraPeers that dramatically increase the scalability.

All that aside, however, centralized schemes like Napster only "scale" in the sense that if you buy another multi-million dollar server and link it to your cluster, you can serve more users. Sure, that's "scaling" in a sense, but it's pretty brain-dead and not particularly interesting from my perspective.

I used to use napster and I never saw above a certain amount of users connected at one time, I believe it did have a ceiling. If I would log off and quickly log back on I am certain I would get another server and all new search results from different clients. The open nap servers are the same all have some max on them and many are full. I used scour mainly untill it was sold. I think it was superior to even morpheus today. Does anyone know how scour was able to search all users? Did they buy more servers?? Or was it software?

11

What do we really know anyway? Most people don't use P2P cause they are afraid to. If they ever get brave, the answer will come!

*sighs* People about "scalability". If one thinks straight, he'll know that the -reach- will always stay around the same number with Gnutella. Why does one have a TTL then? DUH!

For the rest, Gnutella can grow as large as it wants to grow, without adversely affecting itself (abusing clients set aside).

Poor guy...

yea....i'm lost.

all i know is that i could always find what i was looking for with napster, but can't find the same stuff with gnutella. the most likely reason is because those people aren't using gnutella?

(this is a test message)

My question is: who really cares? Gnutella works; it works quitewella. Does it really matter to me if I can't search all quintillion hosts or whatever that are connected? Not a bit.

Trying to discredit Gnutella is like trying to discredit the internet itself. In one form or another, it's here to stay.

I doubt the papers outdated. The Gnutella protocol is fairly basic and probably hasn't changed drastically. In fact if it changed enough to render this document inaccurate it would hardly be the Gnutella network anymore.
You don't need math to prove the Gnutella is about the least efficient P2P network yet. Its just common sense that when you broadcast searches over thousands of nodes it generates a lot of traffic.
However, if it works, it works. If you like it, use it. Its not like Gnutella is going to suddenly shutdown because of some old document.

do you think it'd be a good idea to use the proxy connect function on gnutella apps to make a hyperpeer connect app?
this could connect to many more peers using a limewire-like model,
and relay search and host and data info to any gnutella client.
how easy is this? could some of you programme an emulated proxy app to find and store millions of users and relay searches?
It'd be really FANTASTIC if it happened.

i'm sure most people would agree that gnutella does not work nearly as well as fast-track (think thats the morpheus one that i mean... anyone?). does anyone know why? i believe it must be the way that gnutella is designed. if anyone else around here has read peer-to-peer: harnessing the power of disruptive technologies, then they know that this opinion is patently stolen (although, if your going to steal an opinion, who better to steal it from than o'reilly and associates but, gnutella was designed to find information, not files. it allowed context non-specific searches, and in short was a phat idea. but because there was a lot of people out there keen on sharing mp3s and little else, third-parties wrote software such as the super-node bizniz that allowed searches for files to get more hits. But this removes some of the ingenuety of the protocol, and means that it is being used in a way that it wasn't designed for.

i don't personally see gnutella as the future of file-sharing, and hope that the authors will try to re-invent the system at some point more along its original lines. because of these reasons, i think there are likely to be a number of problems with gnutella, as its not being used for what it was designed for. a protocol like fast-track which is actually DESIGNED with file-sharing in mind is the only way to go, in my (half stolen opinion...

oh, and just for the record, host lists are bad too. those little blighters have been responsible for much of the scalabiluty issues. if everyone would just find an address on a website or by word of mouth like they used to...

oh well. my 2p.