Peer-to-Peer Systems CNT 5517-5564 - PowerPoint Presentation

Slide1

Peer-to-Peer Systems CNT 5517-5564

Dr. Sumi Helal & Dr. Choonhwa LeeComputer & Information Science & Engineering DepartmentUniversity of Florida, Gainesville, FL 32611{helal, chl}@cise.ufl.edu Slide2

The State of the Art of P2P Video Streaming

Slide courtesy: Prof. Darshan Purandare at University of Central Florida, USA Dr

.

Meng

ZHANG,

Dyyno

Inc.,

USA

Jan

David Mol, Delft University of Technology, The

NetherlandsSlide3

IntroductionVideo Streaming

ApproachesIP MulticastContent Distribution NetworkApplication Layer MulticastPeer-to-Peer Swarming ProtocolNoteworthy P2P Streaming SystemsBT-Based ProtocolsCoolStreaming, GridMedia, PPLiveMobile P2P Streaming

OutlineSlide4

P2P Protocols:1999: Napster, End System Multicast (ESM)

2000: Gnutella, eDonkey2001: Kazaa 2002: eMule, BitTorrent2003: Skype2004: Coolstreaming, GridMedia, PPLive

2005~:

TVKoo

,

TVAnts

,

PPStream, SopCast, …Next: VoD, IPTV, Gaming

P2P Is More Than File DownloadSlide5

Internet Traffic

Internet video is ~1/4 of consumer Internet traffic – not including P2PAll forms of video ~90% by 2012TV, VoD, Internet, and P2P

Mobile data traffic will double every year

from now though 2012Slide6

Large-scale video broadcast over InternetReal-time video streamingLarge numbers of viewersAOL Live 8 broadcast peaked at 175,000 (July 2005)

CBS NCAA broadcast peaked at 268,000 (March 2006)NFL Superbowl 2007 had 93 million viewers in the U.S. (Nielsen Media Research)Very high data rateTV quality video encoded with MPEG-4 would require 1.5 Tbps aggregate capacity for 100 million viewersInternet Video StreamingSlide7

IP MulticastContent Distribution Networks

ExpensiveAkamai, Limelight, etcApplication Layer MulticastAlternative to IP MulticastPeer-to-Peer BasedScalableNo setup costViableVideo Streaming ApproachesSlide8

Network layer solutionInternet routers responsible for multicastingGroup membership: remember group members for each multicast session

Multicast routing: route data to membersEfficient bandwidth usageNetwork topology is best known in network layerIP MulticastSlide9

Per-group state in routersHigh complexity, especially in core routersScalability concernViola

tion of the end-to-end design principle: ‘stateless’Slow deploymentChanges at infrastructural levelIP multicast is often disabled in routersDifficult to support higher layer functionalityE.g., error control, flow control, and congestion controlIP MulticastSlide10

CDN nodes deployed at strategic locationsThese nodes cooperate with each other to satisfy an end user’

s requestUser request is forwarded to a nearest CDN node, which has a cached copyQoS improves, as end user receives best possible connectionAkamai, Limelight, etc10Content Distribution Networks (CDNs)Slide11

CDN Example

Origin server (www.foo.com)distributes HTMLreplaces: http://www.foo.com/sports.ruth.gif with http://www.cdn.com/www.foo.com/sports/ruth.gif

HTTP request for

www.foo.com/sports/sports.html

DNS query for www.cdn.com

HTTP request for

www.cdn.com/www.foo.com/sports/ruth.gif

1

2

3

origin server

CDN’s authoritative

DNS server

CDN server near client

client

CDN company

(cdn.com)

distributes gif files

uses its authoritative DNS server to route redirect requestsSlide12

Application layer solutionMulticast functionality in end systemsEnd system

s participate in multicast via an overlay structureOverlay consists of application-layer linksApplication-layer link is a logical link consisting of one or more links in underlying networkMost ALM approaches form tree-based topologyTree construction & maintenanceDisruption in the event of churn and node failures Application Layer Multicast (ALM)Slide13

Easy to deployNo change to network infrastructureProgrammable end hostsOverlay construction algorithms at end hosts can be easily applied

Application-specific customizations13ALM - ProsSlide14

Data-driven/swarming protocolMedia content is broken down in small pieces and disseminated in a swarm

Neighbor nodes use a gossip protocol to exchange their buffer mapNodes trade unavailable piecesBitTorrentP2P Swarming Protocol

CoolStreaming

PPLive

,

SopCast

,

Fiedian

, and

TVAnts

are derivates of

CoolStreaming

Proprietary and working philosophy not published

Reverse engineered and measurement studies releasedSlide15

P2P Swarming Protocol

Pull-based/mesh-basedRedundant chunk avoidanceRobustness and simplicityData availability information rather than an explicit structure to guide data flow (i.e., no need for streaming tree construction)Periodical exchange of data availability with random partners and subsequent retrieval of missing data (i.e., minimal impact from upstream node failures)Higher overhead and longer streaming delay Real-time scheduling constraints (i.e., need for good peer and chunk selection algorithms)Slide16

Tree-Push vs. Mesh-PullSlide17

Tree BasedContent flows from server to nodes along the treeNode failures affect a complete sub-tree

Long recovery timeMesh BasedNodes maintain state information of neighbor nodesResilient to node failureHigh control overheadTree-Push vs. Mesh-PullSlide18

Why Is P2P Streaming Hard?

Real-time constraintsPieces needed in a sequential order and on timeBandwidth constraintsDownload speed >= video speedHigh user expectations

Users spoiled with low start-up time and no/little loss

High churn rate

Robust network topology to minimize churn impact

Fairness difficult to achieve

High bandwidth peers have no incentive to contribute

Slide19

BT-Based P2P Streaming

BitTorrentMeta data (.torrent file)Download policy (piece selection: rarest first)Upload policy (peer selection: Tit-for-tat)Slide20

20

New Download PolicyRequest highest priority piecesHigh prio: download in-order

Mid/low

prio

: download rarest-first

Effect:

dl speed = video speed: peer stays in high

priodl speed > video speed: peer is often in mid/low prioSlide21

BiToS: BitTorrent Streaming

BitTorrent adapted for video streamingChanges to BitTorrent’s piece selection algorithmSlide22

CoolStreamingVideo f

ile is chopped and disseminated in a swarmNode upon arrival obtains a list of 40 peers from the serverNode contacts these peers to join the swarm

E

very

node has typically 4-8 neighbors, periodically sharing its buffer

map

with them

Node exchanges missing chunks with its neighborsDeployed in the Internet and highly successfulSlide23

Membership ManagerMaintains a list of members in the group

Periodically generates membership messagesDistributes it using Scalable Gossip Membership Protocol (SGAM)Partnership ManagerPartners are members that have expected data segments Exchanges Buffer Map (BM) with partnersBuffer Map contains availability information of segmentsSchedulerDetermines which segment should be obtained from which partner

Downloads

segment

s

from

partners and uploads their wanted segmentsCoolStreamingSlide24

Diagram of CoolStreaming SystemSlide25

Designed to support large-scale live video streaming over the InternetThe first generation: Gridmedia I

Mesh-based multi-sender structureCombined with IP multicastFirst release: May 2004The second generation: Gridmedia IIUnstructured overlayPush-pull streaming mechanismFirst release: Jan. 2005GridMediaSlide26

Original GridMediaOverlay construction

Peers self-organize into a richly connected random meshVideo deliveryPeers periodically notifies its neighbor of what packets they hold in the current window of interestEach peer randomly chooses a neighbor to request missing packetsIf a packet does not arrive (i.e., timeout), it is repeatedly requested from a randomly selected neighbor until the packet slides out of the windowPure Random Pull-Based ProtocolSlide27

Hybrid Pull-Push ProtocolPull-based protocol has

trade-off between control overhead and delayTo minimize the delayNode notifies its neighbors of packet arrivals immediatelyNeighbors also request the packet immediately large control overheadTo decrease the overheadNode waits until a group of packets arrive before informing its neighbors Neighbors can also request a

batch of

packets

at a time

considerable delaySlide28

Pull-Push Streaming Mechanism

Pull mechanism as startupSuccessful pulls trigger packet pushes by the neighborsEvery node subscribes to pushing packets from the neighborsLost packets during the push interval are recovered by pull mechanismSlide29

n-sub streams: packets

with sequence number s % nLoop avoidanceFor n-sub streams, there are n packets in a packet groupPacket party is composed of multiple packet groups.Push switching is determined by the pull results of the first packet group in a packet partyPull-Push Streaming MechanismSlide30

Data-driven P2P streamingGossip-based protocolsPeer managementChannel discovery

Very popular P2P IPTV applicationOver 100,000 simultaneous viewers and 40,000 viewers dailyOver 200+ channelsWindows Media Video and Real Video formatPPLiveSlide31

Mobile video streamingRapid growth of mobile P2P communicationVideo streaming expected to rise to as high as 91% of the Internet traffic in 2014Mobile environmentIncrease of mobile and wireless peersUnsteady network connections

Battery powerVarious video coding for mobile devicesFrequent node churnSecurityMobile P2P StreamingSlide32

Mobile P2P StreamingMobile node issuesUplink vs. downlink bandwidthBattery power

Multiple interfacesGeo-targetingOther mobility considerationsProcessing powerLink layer mobilityMobile IP & proxy mobile IPTracker mobilitySlide33

Pioneering ApproachesVideo proxy located at the edge of networksAdaptive video transcoding considering the network conditions and constraints of mobile usersDistributed

transcoding by fixed nodesSub-streams from multiple parents are assembledResilient to peer churnsSlide34

Pioneering ApproachesHierarchical overlayMultiple network interfaces – access link vs. sharing linkPeer fetches a video thru cellular networks (WAN) to share it with others over local networks (LAN)Cooperative video streaming

P2P-based application layer channel bonding in resource-constrained mobile environmentsSimilar, in spirit, to channel/link bundling technology at link layer to efficiently leverage the combined capacity of all access linksSlide35

Questions?