Home → Magazine Archive → March 2023 (Vol. 66, No. 3) → What Ever Happened to Peer-to-Peer Systems? → Full Text

What Ever Happened to Peer-to-Peer Systems?

By Carlos Baquero

Communications of the ACM, Vol. 66 No. 3, Pages 14-15
10.1145/3579633

[article image]

Save PDF

https://bit.ly/3Z9GBzy December 1, 2022

Peer-to-Peer (P2P) systems became famous at the turn of the millennium, mostly due to their support for direct file sharing among users. By the 1980s, the music industry had evolved from selling analogue vinyl records to digital compact disks, but with the introduction of lossy data-compression techniques such as the MP3 coding format, it became feasible to upload/download music files among users' personal computers. Still, content had to be catalogued and found, and P2P systems emerged to provide that functionality.

Some early systems, such as Napster and [email protected], exhibited a mix of P2P and classic client-server architecture. Gnutella and Freenet, the second generation of systems, provided a larger degree of decentralization. The emergence of P2P greatly impacted the business models of the music, and later film, industries. With time, these industries evolved to offer flat rates and subscription services, decreasing incentives for music/video file copying/sharing.

Netflix and other subscription services are a consequence of P2P, but what happened to peer-to-peer as a technological concept?

Looking at Google trends (https://bit.ly/3WPlGAi), we see the concept almost faded from our lexicon. Nevertheless, the technology is used; it evolved and became more specialized. A good portion of the fabric beneath modern datacenters (Web 2.0) and blockchain technology (Web 3.0) evolved from early P2P research. Let's consider some examples.

uf1.jpg
Figure. Google trends for 'Peer-to-Peer'

Search engines predated the rise of P2P and saw significant development in the 1990s. Initially, they were simple inverted indexes that mapped relevant words into document locations on the Web. The first inverted indexes did not try to classify the importance of documents that contained a given word, making difficult the selection of the most relevant choices among results. In the late 1990s, this limitation was removed with the introduction of PageRank, which used the linking structure to rank the importance of pages; Google was created by leveraging that functionality. However, search is easier in a centralized setting.

Gnutella, born in 2000, was one of the first fully decentralized P2P systems, but the search was very primitive, words of interest were propagated to all nodes by flooding, and nodes replied to the source if they had documents matching search terms. There was no ranking. Some users even modified their local node code to pretend they always matched the search terms, to spam other nodes with fake documents. It was becoming clear P2P systems, due to their open and decentralized nature, were particularly subject to attacks.

Around the same time, the Freenet system was introducing additional advances. To avoid document manipulation, it provided Content Hash Keys tied to document content to allow verifying if the returned content matched the key. Another type of key, Keyword Signed Keys, allows for storing content under a readable string, like "The Hobbit," that is transformed by a hash function into a binary key. If different content ended up stored under the same key, the most popular versions would prevail. Both types of keys allowed the lookup of arbitrary data in the network, but efficient lookup was not solved at the time.

Freenet created a topology among nodes that exhibited small-world properties. This guaranteed short paths (with a small number of hops) existed among nodes. However, the routing algorithm was unable to consistently detect those paths with local information, which impacted lookup times. An additional ingredient missing was locality.

The next generation of systems, in the early 2000s, solved this problem by introducing topologies exhibiting locality. The closer one came to the target, the more paths one had to the target and the routing algorithms could pinpoint the next hop with local information and a distance metric. These systems provided users with a Distributed Hash Table (DHT). These efficient content-addressable networks (Chord, CAN, Pastry, Tapestry) allowed structuring N nodes in a topology that supported log(N) routing steps while only storing log(N) network contacts on each node.

Not all these algorithms were easily translated into usable systems. Keeping a structure with nice properties is hard under churn when nodes are failing and being replaced. A subsequent algorithm, Kademlia, added simplicity and robustness to content-addressable networks. It supported symmetric topologies, used a simple XOR distance metric, and allowed concurrent exploration of routes. Variants of Kademlia now are used to support BitTorrent, Ethereum, and IPFS.

The 2000s brought the promise of large-scale P2P systems that aggregated end-user machines and servers. Sadly, these systems tended to have poor quality of service. Portable machines disconnected for the daily commutes and users terminated their P2P processes once their downloads were completed. At the same time, centralized server solutions, typically built on top of SQL databases, also were finding problems keeping up with scaling demand from increasing numbers of Internet users. Both paradigms were lacking in availability. Interestingly, they occupied two extremes in the design space: either full decentralization or full centralization.

The paradigm shifted again. In 2007, Amazon's Dynamo presented a pragmatic system design that built on prior research in DHTs and Eventual Consistency (leveraging file systems research from the early 1990s, Coda and Ficus, which exhibited P2P characteristics before the term was established). In Dynamo the focus was high availability and, unlike prior P2P systems, the nodes were placed under the same administrative control and inside data-centers. The number of nodes scaled down from millions to hundreds, albeit more powerful ones, allowing some simplifications on the DHTs. Availability and low response time were now key concerns, they were good for business.

Quoting from the Dynamo paper: "To meet these stringent latency requirements, it was imperative for us to avoid routing requests through multiple nodes … Dynamo can be characterized as a zero-hop DHT, where each node maintains enough routing information locally to route a request to the appropriate node directly."

The Dynamo design was influential to the NoSQL movement of the 2010s. Several databases were designed following the Dynamo footprints (Project Voldemort, Basho Riak, Apache Cassandra) and featuring the zero-hop DHT approach, but differing in the handling of concurrent updates to replicas. Apache Cassandra is an essential component in modern cloud infrastructure used by Apple, eBay, Instagram, and others.

The phrase "all successful systems attract parasites," often cited in biology, can be applied to P2P systems. Filesharing users terminated nodes once their downloads finished and did not contribute further to the system. Some nodes refused to forward queries from others or lied about uptimes to improve their position in the network.

Different strategies were tested to control free riders: enforcing download and upload quotas to avoid unbalanced downloads; partitioning files in blocks and sharing them in a random order to prevent nodes from quitting when close to having the full file. These strategies tried to coerce users into contributing to the collective, but an important tool was still missing: a clear incentive system.

Hash functions were a key technology in early P2P systems. They provided evidence some content was likely not manipulated. Hashes can be combined in (chain) sequences and direct acyclic graphs to provide more complex certification. In 2008, Bitcoin was presented as "a peer-to-peer distributed timestamp server to generate computational proof of the chronological order of transactions." In these blockchain systems, each peer keeps a local copy of all transactions and can try to augment the chain of transactions by competing/collaborating with other nodes to append a new entry. To prevent Sybil attacks, with one node simulating many other nodes to accrue voting power, these systems can make use of proof of work, spending computation time, or proof of stake, by showing tokens are held.

Blockchains can support the creation of new coins and handle the transfer of coins among users (identified by crypto-graphic key pairs). Coin generation and distribution, particularly when coins can be traded for fiat currency or goods, creates an incentive mechanism to keep the P2P system running.

Blockchain technology is still evolving and seeking winning applications. It might be seen as a continuation of the P2P disruptive revolution 20 years earlier, and its effects on the financial system already are noticeable. We probably need to wait another 20 years to study the legacy of blockchain systems and see which technological concepts turned out to have a lasting impact.

Back to Top

Acknowledgments

I would like to thank Rodrigo Rodrigues, Pedro Henrique Fernandes, and Pedro Souto for their comments on improving this text.

* Further Reading

  1. Peer-to-Peer: Harnessing the Power of Disruptive Technologies. Andy Oram, Ed. O'Reilly and Associates. March 2001.
  2. Rodriques, R., and Druschel, P. Peer-to-peer systems. Commun. ACM 53, 1 (Jan. 2020); https://bit.ly/3X43Dpt

Back to Top

Author

Carlos Baquero is a professor in the Department of Informatics Engineering within the Faculty of Engineering at Portugal's Porto University, and also is affiliated with INESC TEC. His work is focused on distributed systems and algorithms.


©2023 ACM  0001-0782/23/03

Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or fee. Request permission to publish from [email protected] or fax (212) 869-0481.

The Digital Library is published by the Association for Computing Machinery. Copyright © 2023 ACM, Inc.

0 Comments

No entries found