Summary

  • BitTorrent shifted much of a popular file’s outbound traffic from its publisher to people downloading it. It did not remove the web host, tracker or initial seed from the distribution chain.
  • Cohen’s reciprocal choking policy was a client strategy, not a guarantee built into the wire protocol. A peer could contribute more than it received, contribute only enough to improve its own download, or stop sharing.
  • A 2007 study found room for strategic clients to reduce their upload contribution while improving their individual performance. Its results describe the systems and measurements of that study, not every later swarm.

The delivery bill moves with the file

In a conventional HTTP download, the host sends each copy. If a file attracts a sudden audience, the publisher must provision the outbound capacity or accept slower delivery. Bram Cohen’s 2003 paper proposed a different distribution of work: once several people are downloading the same file, they can exchange pieces with one another. The source no longer has to send every byte to every recipient.

That changes where delivery capacity comes from. The publisher’s server still introduces the file, but the swarm’s participants add upload capacity as they arrive. Cohen argued that this could make hosting a very popular file affordable, particularly where a downloader’s upload bandwidth was not metered. That last qualification matters. An unmetered connection can make the transfer feel free to a user, but it does not make the connection, electricity or capacity costless, nor does it describe every access plan.

The design also retained a small but important source role. A publisher placed a .torrent metainfo file on an ordinary web server; a tracker helped downloaders find peers; and a seed with the complete file had to start the distribution. Cohen described the web server and tracker as needing relatively little bandwidth compared with the file transfer, while the seed had to supply at least one complete copy. BitTorrent redistributed the large flow. It did not make the source disappear.

This is why “cheaper hosting” is too broad as a conclusion. The mechanism can reduce the publisher’s outbound burden, but whether it reduces a cash bill depends on bandwidth prices, capacity limits, the audience’s upload availability and the cost of maintaining a seed. It also asks users to supply part of the distribution infrastructure. The economic change is a shift in who contributes delivery capacity, not the elimination of delivery work.

Reciprocity was a local rule, not a promise

Cohen’s paper describes peers deciding which other peers to upload to through a variant of tit-for-tat. A client could temporarily choke a connection, withholding uploads while still allowing downloads; it could also use an optimistic unchoke to try an otherwise unused connection. The trial made room to discover a faster or more useful partner rather than only rewarding an already productive one.

But Cohen explicitly drew a boundary around this mechanism: the choking algorithm was not technically part of the BitTorrent wire protocol. The protocol provided a way for peers to exchange data and signals; a client’s choices about whom to serve were a separate operating policy. That separation matters. It means “BitTorrent guarantees reciprocal exchange” overstates what the shared protocol required. Client behavior, available pieces, connection churn and spare upload capacity all shaped the result.

The publisher’s choice mattered too. Cohen noted that the file owner decided whether to use BitTorrent, while users chose it because it was the route to a wanted file. He also argued that the simple click-to-download interface contributed substantially to adoption and might matter more than the performance and cost-redistribution features. A coordination system can have an elegant incentive rule and still depend on an easy route into the system.

What the 2007 challenge showed

Michael Piatek, Tomas Isdal, Thomas Anderson, Arvind Krishnamurthy and Arun Venkataramani revisited the claim that incentives made BitTorrent robust. In their 2007 USENIX study, modeling based on real-world traces suggested that peers supplied some resources that did not directly improve their own performance. Their BitTyrant client tried to use that gap strategically, adjusting its active peer set and upload rates to obtain more download throughput for less upload.

The conference abstract reports a median 70% performance gain for a 1 Mbit/s client on the live swarms they measured. The authors also warned that universal adoption of strategic clients could reduce average per-swarm performance compared with the client implementations they studied. Those findings challenge the idea that tit-for-tat alone ensured robust contribution. They do not establish what every client or present-day swarm does, and the 70% figure is not a general performance promise.

The distinction is between an incentive that can help organize exchanges and a mechanism that can compel contribution. BitTorrent’s local reciprocity could reward useful uploads, while optimistic trials and uneven capacities left room for contribution that did not pay back directly. A strategic client could try to capture some of that value. Whether widespread strategic behavior would undermine a particular swarm depended on how peers behaved and how much useful capacity remained.

The source remains a point of control

The source’s initial decision still sets the system in motion: publish the metainfo, keep the tracker or another peer-discovery route available, and provide a complete seed. Once the swarm grows, each peer’s client governs its own contribution. No central allocator in Cohen’s design assigns every participant a share of upload work.

That division creates both flexibility and exposure. The publisher can use peers to scale delivery without buying all of the outbound capacity itself. Participants can choose how long to seed or which clients to run. Yet the publisher cannot infer a durable supply of upload capacity from the protocol’s existence alone. If a seed leaves before every piece has spread, or the peers holding a rare piece disconnect, the file may become incomplete. Rarest-first piece selection helps replicate scarce pieces, but it cannot guarantee that a piece remains available after its last holder disappears.

For operators, the useful question is therefore not whether BitTorrent “solved” hosting. It is which traffic the publisher still originates, how much of the load the audience actually absorbs, and what keeps a complete copy available through the tail of demand. Cohen’s contribution was to turn downloaders into potential distributors. The later evidence showed why that capacity should be measured rather than assumed.

Sources