1. P2P Introduction

7/28/2019 1. P2P Introduction

1/52

Eng. Tcnica

Telecomunicacions

Mobile Communication Networks Simon Oechsner - 2013 1

1. Introduction to P2P

Simon Oechsner


2/52

Peer-to-Peer (P2P) removes the strict

separation between clients and servers

Peer-to-Peer



3/52

Peer-to-Peer

Peer-to-Peer (P2P) removes the strict

separation between clients and servers

In P2P, all participants are equal (peers)

and autonomous


server at the same time (servents)

P2P typically also means to implement

application intelligence at the edges of

the network

In the user software, not in separate server

installations


4/52

History of P2P

First Wave P2P Systems (~1999):

Napster: direct peer-to-peer communication, central index server

Gnutella: fully decentralized, use of unstructured overlay networks

Second Wave (2000~2002):


, . ,

eDonkey: distributed index servers, multi-source download

Current Wave (2003~)

BitTorrent: use of tracker (centralized rendezvous nodes), tit-for-tat

Kademlia, Pastry, Chord : fast resource discovery, distributed hash tables

PPLive, Tribler, ...: optimized for video streaming

Skype


5/52

Impact of P2P

Cisco Visual networking index, 2011

Although video streaming is eclipsing P2P file-sharing traffic, itsshare is still significant

The amount oftraffic still grows


Video streamingcan containP2P trafficas well!


6/52

P2P and mobile networks

Two main connections between P2P and mobile communicationnetworks in the context of this course:



7/52

Mobiles/user equipment aspeers




using P2P

Mobile users consumingvideo streams

Content distribution overlays


8/52

P2P as supporting technology

Distribution of centralized



Mobiles/user equipment aspeers


components in the mobilecore network

Mainly useful for databases

Search overlays

using P2P

Mobile users consumingvideo streams

Content distribution overlays


9/52

P2P and scalability

Each new peer adds new resources to the system

Upload bandwidth

Storage capacity

Computing power


,

Each client is also a server

Typically no need for additional infrastructure

Less complexity for deploying systems


10/52

P2P and availability

Peers are generally assumed to be end user hosts

PCs

In the future mobile devices

Thus, peers are assumed to have a low availability and reliability

Can go offline at any point in time


Can crash, data can be corrupted, etc.

Different client versions possible

P2P networks need to consider this and provide mechanisms tocope with very dynamic structures


11/52

Overlays

Logical

overlay

P2P systems create a

logical network

Consisting of the end-to-end

connections between the

peers


Physical

transport

network,

e.g. IP

network

This logical network is called

overlay

To distinguish it from the

underlay, i.e., physical networksubstrate

An overlay is defined on

application layer


12/52

Overlays

Different semantics than the underlying transport network, e.g.neighborhood, address space, and routing

Typically large-scale environment; no administration

Based on locally-made decisions



13/52

Overlays

Different semantics than the underlying transport network, e.g.neighborhood, address space, and routing

Typically large-scale environment; no administration

Based on locally-made decisions


Overlays enable deployment of distributed services without knowing thephysical network

But: blind creation of overlay connections may lead to inefficient usage

of network resources


14/52

Churn

Churn: the continuous process

of nodes joining and leaving

the network

Users may join or leave the

overlay arbitrarily

End user equipment: also


requen no e a ures


15/52

Churn



the network


overlay arbitrarily



Joining

node

requen no e a ures


16/52

Churn



the network


overlay arbitrarily



Joining

node

requen no e a ures

New connections due to new

nodes


17/52

Churn

Failednode



the network


overlay arbitrarily



requen no e a ures


nodes


18/52

Churn



the network


overlay arbitrarily



requen no e a ures


nodes

Re-wiring upon node

failures/leaves

Danger of network

separation

Highly variable topology

New

overlayconnection


19/52

Churn - stochastic scalability

While DHTs are built to scale in the number of nodes,they also have to be able to handle churn

Churn threatens performance and data availability Especially in an enterprise environment, data loss isunacceptable!


Functional Scalability

22 nodes

25 nodes


20/52

Churn - stochastic scalability

While DHTs are built to scale in the number of nodes,they also have to be able to handle churn

Churn threatens performance and data availability Especially in an enterprise environment, data loss isunacceptable!


Functional Scalability Stochastic Scalability

22 nodes

25 nodes

24 nodes,stable stationarystructure

24 nodes,higher churn rate(joins and leaves)


21/52

Self-organization

P2P overlays typically work without a centralizedmanagement authority

or involvement of the user

Global structures emerge and functionality is


Overlay structure repair

Routing of search messages

Data exchange


22/52

P2P functionalities

Resource mediation (search overlays)

search and locate a resource

Performance indicator: search success, search delay/number of routingsteps

Examples: Kazaa, Kademlia, Chord

Resource access and exchange (content distribution overlays)


control, schedule, and conduct exchange of resources or content,

e.g. audio or video files

Performance indicator: download time, stream quality (streaming overlays)

Examples: BitTorrent, eDonkey

Both functionalities can independently be centralized or distributed Index server vs. Search overlay

Download server vs. File-sharing/streaming overlay


23/52

Resource mediation/search overlays

Essentially replace central index servers

Provides information about the location, i.e., IP address, ofdata

Data itself does not need to be stored in the overlay!


and cost considerations Lookup table of several dozen GB in the memory of a server

Often distributed systems provide a higher availability Especially since central servers can be switched off for non-technical reasons


24/52


Search overlays do not have tostore actual user data

No videos, audio, archives, etc.


IP:72.26.203.99

episodev.avi

IP:132.187.16.21


25/52


Search overlays do not have tostore actual user data

No videos, audio, archives, etc.

The important information that isstored is where to find afile/document


This enables the storage (andfinding) of pointerson any node

Not just on the node holding the

actual file

IP:72.26.203.99

episodev.avi

IP:132.187.16.21

Pointer


26/52

Resource access/content distribution overlays

Essentially replace download and streaming servers

Enable the utilization of the normally unused uploadcapacity of end users

Especially powerful if combined with dedicated servers


Scale with the popularity of the file

Handle flash crowdsbetter than servers

Also, offer higher availability and less risk of serviceoutage


27/52

Issues with mobile peers

Mobile peers might want different content than fixed-line peers

Videos in lower resolution

Different formats

r v rl


Who will provide the content? Back to client-server?


28/52

Issues with mobile peers

State-of-the-art CDOs (BitTorrent, eDonkey) reward peers whoupload

Incentive to provide resources

Thus, mobile peers with potentially less upload capacity mightbe at an disadvantage


,

Mobiles in WLANs are not that different from a laptop

Mobile peers may be able to utilize multi-homing

But current clients are typically not prepared for this

WLANAccessPoint

UMTSnodeB


29/52

Classes of P2P networks

-


unstructured


30/52


-


unstructured

Hybrid P2P


31/52


-


unstructured

Hybrid P2P Hierarchical P2P


32/52


- -


unstructured

structured

Hybrid P2P Hierarchical P2P


33/52

Unstructured P2P

Simplest form of an overlay

Random neighbor selection Mesh topology


Just establish new connections with any potential neighbor

May show small-world characteristics

Any node can be reached from any other node with a smallnumber of (overlay) hops


34/52

Unstructured P2P - Resilience

If all nodes are really equal, an unstructured P2P network is veryresilient against random node failures

However, in reality, nodes are not equal Different amounts of resources

Different uptime



35/52

Unstructured P2P - Resilience

If all nodes are really equal, an unstructured P2P network is veryresilient against random node failures

However, in reality, nodes are not equal Different amounts of resources

Different uptime


This may lead to preferential attachment of new nodes

the rich get richer

Power-law distribution for the number of connections of individualpeers

If the nodes with a high number of connections fail, the networkcan become separated


36/52

Unstructured P2P - Search

Problem: which nodeholds the data/entry/pointer

searched for?

No structure: any node may

??


Or none!

No informed search possible

Easiest solution: flooding

Fl di


37/52

Flooding

Each node forwards received queries to all its neighbors

Minus the node from which the query was received


Fl di


38/52



Flooding


Fl di


39/52

Flooding




up ca e quer es

received!

Fl di


40/52

Flooding



Inefficient network resource usage Typically, only a few nodes hold the searched information

In the worst case, every node is queried


Duplicate queries due to random graph structure

Overhead can be limited

Lifetime of queries (TTL in hops)

If not all nodes are queried, search may be unsuccessfulalthough the searched information exists!

R d m lk h


41/52

Random walk search

Instead of fowarding a query to all neighbors, onlyone is chosen randomly

Or a small number k

Limits the number of queries


But also limits further the success probability of thesearch within a given time

Alternatives for searching


42/52


Hierarchical overlays

Higher layers compile theinformation from all nodesof lower layers


Queries are routed tohigher layers with lessnodes

There, flooding or othersearch algorithms can takeplace



43/52


Breaks up the concept of all peers being equal

Some peers get more responsibility super peers

Can be selected based on their characteristics


Peers on each level or layer can form an additionalnetwork

Mobile peers & hierarchical overlays


44/52

Mobile peers & hierarchical overlays

A hierarchical overlay might be a good choice to integratemobile peers

Especially when they cannot guarantee availability all the time

Fixed-line superpeers can handle the more critical functions

Search


Possibly dedicated servers as proxies for mobiles

Mobile peers only get the results

Less overhead due to (mobile) churn

Example Application: Skype


45/52

Example Application: Skype

Skype is at its core still P2P VoIP, while offering additionalservices

E.g., SkypeIn/SkypeOut, video conferencing,

The Skype protocol is proprietary, but measurement studieshave been conducted


Skype uses its participating nodes to store user data

Contact list, status, preferences, IP

Skype uses a hierarchical search and signaling overlay

Ordinary nodes are connected to supernodes

Supernodes are elected based on performance andonline time



46/52


Structured overlays

Information is stored on specific nodes

according to well-known rules Remember: pointers can be stored on any node!


This enables queries for an information to be routed to the node

containing it (key-based routing)

Routing on overlay level

Much higher guarantee of search sucess With less overhead in terms of search messages

but higher overhead to keep the overlay structure!

Internal overlay traffic


47/52


Since overlays are a distributed system, most queries to thissystem are forwarded and routed internally

(mainly true for search overlays)




48/52




With each internal hop, the load on the system increases

more bandwidth is needed




49/52




With each internal hop, the load on the system increases

more bandwidth is needed


The real load on the system is the sum of the external load andinternal load

dimensioning is based on this real load

for the system as a black box, only the external load is aperformance indicator

Internal load increase in overlays


50/52

Internal load increase in overlays

0 = 2m

IM 2


Initial Query

InternalMessage 1

IM 3

Factorinternal load/external load

= 3!

Traffic overhead vs mobile networks


51/52

Traffic overhead vs. mobile networks

Each additional hop via a mobile peer means usage of the (incomparison) costly air interface

Therefore, the less hops are taken via mobile peers, the moreresource-efficient an overlay is




52/52


Each additional hop via a mobile peer means usage of the (incomparison) costly air interface

Therefore, the less hops are taken via mobile peers, the moreresource-efficient an overlay is


It might even be better to not use the upload capacity ofmobiles/wireless devices in content distribution networks

Might not contribute much anyways

Upload might compete with download!

Documents

1. P2P Introduction