On Thursday, I jumped in to a bloggic discussion of the tradeoffs between centrally-controlled and peer-to-peer design strategies in distributed systems. (See posts by Randy Picker (with comments from Tim Wu and others), Lior Strahilevitz, me, and Randy Picker again.)

We’ve agreed, I think, that large-scale online services will be designed as distributed systems, and the basic design choice is between a centrally-controlled design, where most of the work is done by machines owned by a single entity, and a peer-to-peer design, where most of the work is done by end users’ machines. Google is a typical centrally-controlled design. BitTorrent is a typical P2P design.

The question in play at this point is when the P2P design strategy has a legitimate justification. Which justifications are “legitimate”? This is a deep question in general, but for our purposes it’s enough to say that improving technical or economic efficiency is a legitimate justification, but frustrating enforcement of copyright is not. Actions that have legitimate justifications may also have harmful side-effects. For now I’ll leave aside the question of how to account for such side-effects, focusing instead on the more basic question of when there is a legitimate justification at all.

Which design is more efficient? Compared to central control, P2P has both disadvantages and advantages. The main disadvantage is that in a P2P design, the computers participating in the system are owned by people who have differing incentives, so they cannot necessarily be trusted to work toward the common good of the system. For example, users may disconnect their machines when they’re not using the system, or they may “leech” off the system by using the services of others but refusing to provide services. It’s generally harder to design a protocol when you don’t trust the participants to play by the protocol’s rules.

On the other hand, P2P designs have three main efficiency advantages. First, they use cheaper resources. Users pay about the same price per unit of computing and storage as a central provider would pay. But the users’ machines a sunk cost – they’re already bought and paid for, and they’re mostly sitting idle. The incremental cost of assigning work to one of these machines is nearly zero. But in a centrally controlled system, new machines must be bought, and reserved for use in providing the service.

Second, P2P deals more efficiently with fluctuations in workload. The traffic in an online system varies a lot, and sometimes unpredictably. If you’re building a centrally-controlled system, you have to make sure that extra resources are available to handle surges in traffic; and that costs money. P2P, on the other hand, has the useful property that whenever you have more users, you have more users’ computers (and network connections) to put to work. The system’s capacity grows automatically whenever more capacity is needed, so you don’t have to pay extra for surge-handling capacity.

Third, P2P allows users to subsidize the cost of running the system, by having their computers do some of the work. In theory, users could subsidize a centrally-controlled system by paying money to the system operator. But in practice, monetary transfers can bring significant transaction costs. It can be cheaper for users to provide the subsidy in the form of computing cycles than in the form of cash. (A full discussion of this transaction cost issue would require more space – maybe I’ll blog about it someday – but it should be clear that P2P can reduce transaction costs at least sometimes.)

Of course, this doesn’t prove that P2P is always better, or that any particular P2P design in use today is motivated only by efficiency considerations. What it does show, I think, is that the relative efficiency of centrally-controlled and P2P designs is a complex and case-specific question, so that P2P designs should not be reflexively labeled as illegitimate.

There’s an conversation among Randy Picker, Tim Wu, and Lior Strahilevitz over the U. Chicago Law School Blog about the relative merits of centralized and peer-to-peer designs for file distribution. (Picker post with Wu comments; Strahilevitz post) Picker started the discussion by noting that photo sharing sites like Flickr use a centralized design, rather than peer-to-peer. He questioned whether P2P design made sense, except as a way to dodge copyright enforcement. Wu pointed out that P2P designs can distribute large files more efficiently, as in BitTorrent. Strahilevitz pointed out that P2P designs resist censorship more effectively than centralized ones.

There’s a subtlety hiding here, and in most cases where people compare centralized services to distributed ones: from a technology standpoint, the “centralized” designs aren’t really centralized.

A standard example is Google. It’s presented to users as a single website, but if you look under the hood you’ll see that it’s really implemented by a network of hundreds of thousands of computers, distributed in data centers around the world. If you direct your browser to www.google.com, and I direct my browser to the same URL, we’ll almost certainly interact with entirely different sets of computers. The unitary appearance of the Google site is an illusion maintained by technical trickery.

The same is almost certainly true of Flickr, though on a smaller scale. Any big service will have to use a distributed architecture of some sort.

So what distinguishes “centralized” sites from P2P designs? I see two main differences.

(1) In a “centralized” site, all of the nodes in the distributed system are controlled by the same entity; in a P2P design, most nodes are controlled by end users. There is a technical tradeoff here. Centralized control offers some advantages, but they sacrifice the potential scalability that can come from enlisting the multitude of end user machines. (Users own most of the machines in the world, and those machines are idle most of the time – that’s a big untapped resource.) Depending on the specific application, one strategy or the other might offer better reliability.

(2) In a “centralized” site, the system interacts with the user through browser technologies; in a P2P design, the user downloads a program that offers a more customized user interface. There is another technical tradeoff here. Browsers are standardized and visiting a website is less risky for the user than downloading software, but a custom user interface sometimes serves users better.

The Wu and Strahilevitz argument focused on the first difference, which does seem the more important one these days. The bottom line, I think, is that P2P-style designs that involve end users’ machines make the most sense when scalability is at a premium, or when such designs are more robust.

But it’s important to remember that the issue isn’t whether the services uses lots of distributed computers. The issue is who controls those computers.