CD-GAIN: Content DeliveryThrough the Analysis of Users' Access Patterns

Dmytro Karamshuk1, Nishanth Sastry1, Jigna Chandaria2, Andy Secker2

1King's College London, 2BBC R&D

CD-GAIN: ESPRC project in collaboration with BBC R&D, http://bit.ly/cd-gain

CISCO Visual Networking IndexSome Figures

• Video traffic in 2013 accounted for 66% of all trafficon the Internet and is predicted to reach 79% by 2018

• Nearly three-fourths of the world’s mobile data traffic will be video by 2019

…which Raise Some IssuesUser Experience Carbon Footprint

Business Models

Users - 32 M/month

IP address – 20 M/month

Sessions - 1.9 Billion

May 2013 – Jan 2014

≈ 50% of the UK population

First-mile view of the UK’s IP-ecosystem

2 x [INFOCOM’2015], [WWW’2013]

Where the market is heading?Fixed-line Internet market (5 representative providers)

Mobile market is more dynamic than the fixed-line Internet market

Mobile Internet market (5 representative providers)

Pricing Models

All-you-can-eat data(M1, M5)

Limited-cap data packages(M2 – M4)

All-you-can-eat plans boost users consumption

Users of all-you-can-eat data providers are more “loyal”

Behavioral Shift• Fixed-line accesses (F1-F5) peaks in the evening

• Limited data caps (M2-M4) peak in commute hours

Mobile users watch more during commutes

• Unlimited plans (M1, M5) is a superposition of both

Product BundlesInter-ISP attractiveness: normalized fraction of shared users

between a pair of mobile and fixed-line ISP

Product bundles can be used to share infrastructure between peaking in non-intersecting time periods mobile and fixed-line ISPs

Combinations of ISPs corresponding to existing

product bundles

Limitations to mobile-fixed interactionsCorrelation of content

popularity across different ISPsShare of accesses per content genre

User of mobile and fixed-line ISPs watch slightly different content

More on Infrastructural SupportRegional activity levels are correlated with the access speed

Solution: let’s install more distributed caches!!!

Content Delivery: traditional approach

Install more distributed caches Requires investments

Increased carbon footprint

Peer-assisted Content Delivery

Content Delivery Network

user

user user

user

useruser

average of 5K users online every sec in the first day after release

5K duplicates

Ask users for assistance

P2P: ObstaclesBitrate Stratification

Partial Participation

ISP friendliness1080p

720p

480p

Dealing with ObstaclesFive biggest ISPs account for over

80% of trafficTwo main bitrate formats

dominate in over 70% of sessions

Traffic fragmentation by ISP and bitrate formats yields few large content swarms

Traffic Savings and Swarm CapacityTraffic gain:

- volume of traffic downloaded from content delivery nodes

- total volume of traffic watched by the users

potential traffic gains for popular content are immense despite constraints

Swarm capacity: c

G 1Ts

Tu- average number of users online

A Simple Theoretical Model

model the system as an infinite servers queue

CDN user

1

user

2

user

3

intuition: the more users in a swarm the more savings

Higher Demand means Higher Savingsci=ui r iAverage number of peers online

G(c)=1−(

c(1+ec c−m(mΓ(m)−Γ(1+m,c)))

m+1)

−1

How much we can offload to peers

Traffic gains are growing with the demand and the size of content

request ratemean session duration

G c 1 e c

Particularly, if m=1:

Theory VS Simulations

Theory is in a good agreement with simulations

Savings across Content CorpusTop-5% of the content

corpus accounts for 80% of trafficMost of accesses happen in the

first day after release

Yes, it’s all about very popular content

https://vimeo.com/99054874

How about Unpopularity Content?

Content Bundling Client Caching

Pro-active pushing VS reactive pulling?

Client Caching

client caching can boost gains for up to 20%

Improv. in a single swarm Improv. for the content corpus

Carbon Footprint

number of hops = 5-7 number of hops = 3

Traditional content delivery network Peer-assisted CDN

Number of networking nodes powered to delivery a video in peer-assisted scenario is potentially smaller

Potential Energy SavingsAverage distance between peers

in a swarm with a given capacityEnergy savings for a swarm

with a given capacity

peer-assisted content distribution may lead to energy savings

Conclusions• there is substantial difference in accesses patterns between fixed-line ISPs, mobile ISPs with all-you-can-eat data and limited data-caps

• this difference can be effectively exploited in shared mobile-fixedinfrastructures

•user engagement depends on the level of infrastructural support

• with peer-assisted delivery savings grow with demands evendespite obstacles

• peer-assisted CDN can also be green

Dmytro KaramshukKing's College London

follow me on Twitter: @karamshuk

Our website: http://bit.ly/cd-gain