P2P (Peer-to-Peer) Network

Motivation

The goal of a P2P network is to offload the burden on the service provider by transfering computational, storage bandwidth resource responsibilities to the clients. It is a service architecture used to provide highly scalable & cost efficient services.

Application

A P2P network can be used for:

  • Content Distribution: Napster, BitTorrent
  • Internet Telephony: Skype
  • Video Streaming: WebRTC
  • Computation: Bitcoin

Architecture

Centralized Lookup Service

If I have some content that I want to share, how do I let everybody else on the network know that it exists? Or if I want to search content on the network, how do I know who has it?

A centralized lookup service is a directory that keeps tracks of who has what content and how it can be contacted. In the example below:

  1. Machine m1 is looking for content E, but it doesn’t know which machine on the network has it.
  2. m1 contacts the centralized lookup service, and is informed that machine m5 has the content.
  3. m1 then connects with m5 directly using the its IP & port as provided by the centralized lookup service to get content E.

Image from UC Berkeley, CS 162, Fall 2011

Pros:

  • Easy to implement & deploy sophisticated features on top of a single service

Cons:

  • Service needs to scale to handle growing nodes participating on the network
  • Single point of failure & single source of legal vulnerability

Gnutella

One way to solve the single point of failure problem with centralized lookup service is to de-centralize the lookup. But now having removed the lookup service, how would I know where to get content?

When I join the network as a node, I will pick randomly from a list of pre-defined peers (typically hardcoded in the application) and join their network. These peers & their immediate neighbors now become my neighbors. I can then send my requests to my neighbors & have them forward my request downstream. In the example below:

  1. Machine m1 is looking for content E, but it doesn’t know which machine on the network has it.
  2. m1 sends a quest to its neighbors m2 & m3.
  3. Neither m2 or m3 have the content, so they forward the request to their neighbors m4, m5 and m6.
  4. m5 happens to have the content, so it contacts m1 to exchange the content.

Image from UC Berkeley, CS 162, Fall 2011

Note:

A request contains the requesting node’s IP & Time-To-Live value - if a request’s Time-To-Live reaches 0 before the content is found, the request is dropped altogether. This limits the possibility of requests flooding the network.

Pros:

  • Completely decentralized

Cons:

  • Requests can potentially flood the network as nodes grow, ending up with more query traffic than content traffic
  • Horizon effect introduced by Time-To-Live; common resources are easy to find while rare resources can seldom be reached

KaZaa

In addition to some of the problems seen in the previous topologies, there are other challenges yet to be addressed in P2P networks:

  • Dynamicity: Each node on the network can be online & offline at any time. Requests might not make it too far downstream before it is dropped.
  • Heterogeneity: Each node on the network has different capacity (bandwidth, computation power). Requests going through a deprived node could lead to a performance bottleneck.

To solve these problems we identify the nodes that are consistently online with good capacity, and designate them as ultra-peers. We can assign these ultra-peers as the first point of contact to all other leaf nodes. Utlra-peers themselves are also connected to fellow ultra-peers, forming a two-level topology. A request going through an ultra-peer is advertised to all leaf nodes & other ultra-peers.

Image from UC Berkeley, CS 162, Fall 2011