1 / 12

Design considerations for P2P

Design considerations for P2P. Stefan Saroiu, Krishna Gummadi, Steve Gribble Department of Computer Science and Engineering University of Washington. {tzoompy,gummadi,gribble}@cs.washington.edu. What this talk is about. Peer-to-Peer systems have arrived but, are they here to stay?

Faraday
Télécharger la présentation

Design considerations for P2P

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Design considerations for P2P Stefan Saroiu, Krishna Gummadi, Steve Gribble Department of Computer Science and Engineering University of Washington {tzoompy,gummadi,gribble}@cs.washington.edu

  2. What this talk is about • Peer-to-Peer systems have arrived • but, are they here to stay? • many things work well, but many things are still missing • Goals of this talk: • to argue that current P2P designs are missing crucial properties • to show how specific architectural decisions have led to the loss of these properties • to encourage this community to recapture these properties in their designs, so that our systems are practical and successful Saroiu, Gummadi, Gribble

  3. The Good and the Bad • P2P systems have achieved several goals • decentralized and symmetric infrastructure • evenly sharing resources and responsibilities among participants • But, some things are still missing • we’ve lost a few “ilities”: • securability: more vulnerable to malicious attack • composability: can’t transparently compose functionality into the system • predictability and controllability: cannot easily engineer system to provision for performance, reliability, workload or value Saroiu, Gummadi, Gribble

  4. How to show why this has happened • Start with a well understood distributed system that is known to have all of the “ilities” • the WWW • Compare its architecture to a representative P2P system • Chord/CFS [but we could use any of CAN, Pastry, Tapestry, …] • Figure out the differences, and draw implications • show how the difference lead to loss of the properties • Convince you that we need to recapture back these “ilities” Saroiu, Gummadi, Gribble

  5. The Architecture of the Web

  6. Contrasting the Web with P2P

  7. Collapsing name and address space • A name is an address • the ability to create an address implies control over certain names • lose controllability: cannot separately grant authority over names and addresses • must engineer other (higher level) mechanisms • The name of content dictates the node that must manage it • inserting content into the system forces others to do work on my behalf, and others can force me to do work • lose securability: denial of service attacks are possible Saroiu, Gummadi, Gribble

  8. Collapsing routing and lookup • Name-based routing • lack of explicit lookup system removes a level of indirection • lose composability: harder to make replication transparent to the client • Servers must now be routers: roles are not separable • cannot have different levels of trust • Web: core routers typically more trustworthy than web servers • P2P: lose controllability, cannot engineer trust according to role • failure of routing and content serving is intertwined • Web: failure of server doesn’t affect routing to other servers • P2P: lose controllability, isolation of failures is lost Saroiu, Gummadi, Gribble

  9. Topology is deterministic • Chord IDs of participants dictates overlay topology • lose securability: possible to attack system by choosing content/server address • can hijack content by choosing appropriate address • can surround and “monitor” a node in system by choosing address • lose controllability: harder to do policy-based routing • set of routes available is dictated by nodes that participate Saroiu, Gummadi, Gribble

  10. Topology is randomized • Hashing spreads mapping of overlay to physical links • lose controllability and predictability: can’t predict the nodes or physical links involved with content • cannot do local provisioning for hotspots, high value content • content provider must trust strangers to provide high quality of service • randomness amplifies bad local properties of system, but not good • a 1000 node Chord overlay with 20% modems has >80% slow paths Saroiu, Gummadi, Gribble

  11. Summary • Architectural decisions have several implications • helped to simplify systems and achieve many good properties • but, have also lost several crucial “ilities” • composability: transparently compose functionality into system • securability: prevent certain classes of attacks • e.g., denial of service by forcing others to do work • predictability: anticipate which nodes/links are involved with content • randomness diffuses local properties globally, and amplifies the bad • controllability: ability to engineer roles, resources, and responsibilities • e.g., don’t want to have to depend on others for quality of my content Saroiu, Gummadi, Gribble

  12. Moving Forward • To mature P2P systems into a practical application infrastructure, we need to recapture the “ilities” • ability to enforce: • who publishes, who participates • ability to engineer according to: • specific load, value of content • ability to delegate and engineer responsibilities by: • trustworthiness, capability of contributing resources • Otherwise, our systems will fail for reasons that are incidental to the current design goals • why? because these are the pragmatic, engineering reasons that world needs to make a system successful in practice Saroiu, Gummadi, Gribble

More Related