Peer-to-Peer Computing

CS587x LectureDepartment of Computer Science

Iowa State University

What to Cover

Review on some P2P applications Napster Gnutella Freenet

Discussion and summary

Resource Sharing

Questions to answer in order to design a resource-sharing network How to add new nodes to the network How can one node know about others How can a node find and retrieve data How to manage the shared data

users

Client/Server ArchitectureCreate a server to store the information that these nodes want to share

The server is the only data source Clients request data from serverExample: mp3.com

A client registers to mp3.com and uploads its music files to the server

The songs are then stored and indexed on a server that is part of the web site

Other uses can connect to the web site and downloads the songs they are interested in

Limitation of C/S model Scalability is hard to achieve Presents a single point of failure Requires administration Unused resource at the network edge

Some P2P Applications

NapsterGnutellaFreenet

NapsterEach node registers to napster.com and provides a list of its song titles

The napster server knows the music titles and their sites

The songs themselves are still stored locally

For a node to download a song,

the node contacts the server

The server returns a list of nodes that have the song

The requesting node selects one of the nodes in the list to download the file directly from the node

Highlights of Napster

Main innovation: a client downloads a music directly from another client, i.e., P2P communication

After a client downloads a music, it can serves other clientsNapster server itself does not have any music files

It acts as a directory or brokerAdvantages

Each consumer contributes its resource (disk and bandwidth) and content to the community

Contents are more reliable because the same file is stored in many nodes, which are geographically distributed

Administration and service cost are minimalDrawback

Napster is a hybrid P2P system since a central server is required to coordinate file sharing

The central server presents a single point of failure

Gnutella

Creating a Gnutella network A node joins the network with a PING to

announce self IP address, port, number/size of shared files

Receivers forward the Ping to their neighbors Receivers back-propagate a PONG to announce

self Each Pong includes sender’s IP address, number/size of

shared files

Maintaining a Gnutella network PING neighbors periodically PING Well-known root nodes if starting from

scratch

Search Protocol

For node A to request a file (any kind), itcreates a query (A, S, N, T), where S is search string, N unique request ID, T Time-to-Livechecks local system, if not found

Sends (A, S, N, T) to all Gnutella neighbors

B receives a query (A, S, N, T) If B has already received query N or T = 0, drops the

query Otherwise, B looks up S locally and sends (N, Result) to A

if anything found Any kind of look up (could simply grep, or construct some sql

cmd) If not found locally,

B sends (B, S, N, T-1) to all of its Gnutella neighbors B records the fact that A has made the request N

When B receives a response of the form (N, Result) from one of its neighbors, it forwards the response to A

Gnutella Messages

PING request the transitive closure of connected nodes to identify

them, essentially asking the question "Are you there?“PONG

response by a node upon receiving a PING; the responding node provides its IP address and number of sharable files it contains. This gives the answer that "Yes, I am here….“

QUERY request to locate a set of files matching some filter criteria.

These are messages stating, "I am looking for x".HITS

response to a query giving a list of files matching the filter criteria and the IP address of the provider, can be many in number.

GET/PUSH request a file provider to contact the requester. This

provides a simple mechanism to attempt to get through firewalls

Partial Map of a Gnutella Network

Highlights of Gnutella

Pure P2P Unlike Napster, Fully decentralized, no single

point of failure

Limitations Scalability: if you send out a request with a

TTL of 10, and each site contacts six other sites, up to 61+62+63+64+65 +66+67+68+69+610 messages could be exchanged

Not anonymous: since result contains the URL string, the source provider can be tracked – this is addressed in Freenet

Freenet

Freenet is a pure P2P system mainly designed to support distributed information storage and retrieval anonymity for producers, consumers and

holders of information adaptive respond to usage patterns

Freenet differentiates from Gnutella mainly in Retrieving data Storing data Managing data

ArchitectureEach file is identified by a binary key The key is generated using some hash function Every file is stored, retrieved, and maintained with

its file key

Each node maintains a local data store and a routing table data store maintains a set of files routing table keeps information about neighboring

nodes and the keys that they are thought to hold A sequence of (file key, node address) Used for file retrieval

key neighbor

30 123.234.456.1

100 888.234.456.2

65 999.234.456.3

Retrieving dataA user first obtains or calculates a keyThe user sends a search request message (key+TTL) to local nodeWhen a node receives a request, it checks its own data storage

If the specified data is found, returns it Otherwise, the node looks up its routing table and

forwards the request to the node that has the nearest key

why do this - the similarity of two keys actually has nothing to do with that of their corresponding files?

If this request is successful, the node that has the target data

returns the data through the search path, caches the file in its own data store, and creates a new entry in its routing table

key neighbor

30 123.234.456.1

100 888.234.456.2

65 999.234.456.3

Example

A B

C

1. Check datastore for file

Data reply + actual data source

FOUND

DE

NOT FOUND

2. Check routing table for node with nearest key to requested one

File request (key, hops to live)

Failure message

Cache file in datastoreCreate new entry in routing table

Cache file in datastoreCreate new entry in routing table

Cache file in datastoreCreate new entry in routing table

NOT FOUND

FAILURE

3. Try the node with second nearest key

1. Calculate binary file key2. Check routing table for node with nearest key

Effect of Retrieving Mechanism

Anonymity Uncontrolled replication allows one to deny

responsibility of having the fileQuality of routing improved over time: Nodes specialize in locating sets of similar keys Files with similar keys are stored in clustering

(why?) Files are key-clustering instead of subject-clustering

Transparent replication of popular data Improved data availability Replication degree depends on data popularity

Increasing connectivity The graph becomes more and more connected

Effect of Retrieving Mechanism

Major difference from Gnutella searching Breadth-first search vs. Depth-first

search Replication over the retrieval path

Limitation Searching for a document that does

not exist?

Storing data

Calculate binary file key and send insert message like request (key, hops to live)When a node receives an insert proposal, it first checks its own data store

If the key already exists, the users need to try again using different key

Otherwise, the node looks up the nearest key in its routing table and forwards the insert to the corresponding node

If key collision occurs at the adjacent node, the node notifies the inserted to try another key

If TTL expires without a key collision, an “all clear” result will be backwarded to the original inserter

Storing data

Effects of insert mechanism: New files are placed on nodes possessing

files with similar keys

Limitation How long it takes to insert a file? How about version management? Two different files could have the same key

and both may exist in network Different users must have different name space The same user must use different file description

(e.g., keywords) for different file Security is a concern

Managing data

File replacement is done using LRU Data items sorted in decreasing order by time

of most recent request/insert Outdated documents fade away naturally as

routing table entry will remain for a time

File lifetime The time period of keep a file is unknown You cannot delete a file from a Freenet – a file

will not disappear unless it is not accessed for a while

No guarantee that a document you submit today will exist tomorrow

Highlights of Freenet

Pure P2P - similar to Gnutella, Provides anonymity

Neither data producer and retriever can be identified

Searching/Storing/Managing are all different for anonymity and performance purpose

P2P Advantages

Efficient use of resources Client/Server architecture cannot take advantage of the

unused bandwidth, storage, processing power at the edge of network

Scalability Each user contributes its resource to the entire

community, instead of just a burden

Reliability Replicas Geographic distribution No single point of failure

Ease of Administration Nodes self organize No need to deploy servers to satisfy demand Built-in fault tolerance, replication, and load balancing

P2P Computing SummaryP2P computing is the sharing of computer resources by direct exchange between systems

Such resource includes information, processing cycles, storage, etc.

A P2P network has the following characteristics Each node behaves as client, server, and router Nodes are autonomous (no administrative authority) Network is dynamic: nodes enter and leave the

network frequently Nodes collaborate directly with each other (not

through well-known servers) Nodes have widely varying capabilities

Homework 3 (Due 04/20)

Implement a Gnutella network Network maintenance (60/100 points)

ping and pong Nodes being able to retrieve files

(40/100) query, hit, get