Akamai recently announced that its infrastructure will “have to expand by a factor of 100 times in the next five years just to keep up with the demand for real-time video.” One of the reasons comes from the rate-adaptive streaming technologies. Our mission is to reduce the footprint of live rate-adaptive streaming applications on the CDN infrastructure. We show in this presentation that a smart system can reduce the infrastructure needs by a factor of five with negligible losses of Quality of Experience (QoE) for end users.

1. Scadoosh:
Scaling Down CDN
Infrastructure
Gwendal Simon

2. Motivations
Akamai’s infrastructure will have to expand by a factor
of 100 times in the next five years just to keep up with
the demand for real-time video.
Paul Sagan, Akamai CEO, Jun 2012
2 / 15 Gwendal Simon Scadoosh

3. Live stream delivery
Content Provider CDN
edge
encoders Clients
servers
ingest origin
server server
3 / 15 Gwendal Simon Scadoosh

4. Live stream delivery
Content Provider CDN
edge
encoders Clients
servers
ingest origin
server server
disclaimer : ISP strategy is out of the scope of this talk
3 / 15 Gwendal Simon Scadoosh

5. Rate-adaptive video streaming
HTTP streaming server client
video segments
requests
4 / 15 Gwendal Simon Scadoosh

6. Rate-adaptive video streaming
HTTP streaming server client
video segments
requests
representation 3
representation 2
representation 1
time
4 / 15 Gwendal Simon Scadoosh

7. Rate-adaptive video streaming
HTTP streaming server client
video segments
requests
bit-rate
representation 3
available bandwidth
representation 2
representation 1
time time
4 / 15 Gwendal Simon Scadoosh

8. Rate-adaptive video streaming
A1 50 kbps 320 x 240 D1 900 kbps 1280 x 720
A2 100 kbps 320 x 240 D2 1.2 Mbps 1280 x 720
A3 150 kbps 320 x 240 D3 1.5 Mbps 1280 x 720
B1 200 kbps 480 x 360 D4 2.0 Mbps 1280 x 720
B2 250 kbps 480 x 360 E1 2.5 Mbps 1920 x 1080
B3 300 kbps 480 x 360 E2 3.0 Mbps 1920 x 1080
B4 400 kbps 480 x 360 E3 4.0 Mbps 1920 x 1080
B5 500 kbps 480 x 360 E4 5.0 Mbps 1920 x 1080
C1 600 kbps 854 x 480 E5 6.0 Mbps 1920 x 1080
C2 700 kbps 854 x 480 E6 8.0 Mbps 1920 x 1080
5 / 15 Gwendal Simon Scadoosh

9. Rate-adaptive video streaming
A1 50 kbps 320 x 240 D1 900 kbps 1280 x 720
A2 100 kbps 320 x 240 D2 1.2 Mbps 1280 x 720
A3 150 kbps 320 x 240 D3 1.5 Mbps 1280 x 720
B1 200 kbps 480 x 360 D4 2.0 Mbps 1280 x 720
B2 250 kbps 480 x 360 E1 2.5 Mbps 1920 x 1080
B3 300 kbps 480 x 360 E2 3.0 Mbps 1920 x 1080
B4 400 kbps 480 x 360 E3 4.0 Mbps 1920 x 1080
B5 500 kbps 480 x 360 E4 5.0 Mbps 1920 x 1080
C1 600 kbps 854 x 480 E5 6.0 Mbps 1920 x 1080
C2 700 kbps 854 x 480 E6 8.0 Mbps 1920 x 1080
one video stream = a dozen of representations
= more than 20 Mbps
5 / 15 Gwendal Simon Scadoosh

10. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors
edge servers
ISP 1 ISP 2 ISP 3
6 / 15 Gwendal Simon Scadoosh

11. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors
edge servers last-mile ?
ISP 1 ISP 2 ISP 3
6 / 15 Gwendal Simon Scadoosh

12. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors
edge servers last-mile ?
ISP 1 ISP 2 ISP 3
adaptive streaming → last-mile
6 / 15 Gwendal Simon Scadoosh

13. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors
peering link ?
edge servers
ISP 1 ISP 2 ISP 3
adaptive streaming → last-mile
6 / 15 Gwendal Simon Scadoosh

14. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors
peering link ?
edge servers
ISP 1 ISP 2 ISP 3
adaptive streaming → last-mile
edge servers in ISP → peering link
6 / 15 Gwendal Simon Scadoosh

15. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors in edge-servers ?
edge servers
ISP 1 ISP 2 ISP 3
adaptive streaming → last-mile
edge servers in ISP → peering link
6 / 15 Gwendal Simon Scadoosh

16. Identifying the issue
Where is the bottleneck in today’s CDN ?
origin servers
reflectors in edge-servers ?
edge servers
ISP 1 ISP 2 ISP 3
adaptive streaming → last-mile
edge servers in ISP → peering link
commoditized hardware→ server under-provisioning
6 / 15 Gwendal Simon Scadoosh

17. Our objective
Finding a trade-off between :
infrastructure capacity for CDN providers
quality of experience for clients
7 / 15 Gwendal Simon Scadoosh

18. Our objective
Finding a trade-off between :
infrastructure capacity for CDN providers
quality of experience for clients
Typical services : multiple simultaneous live streams
user-generated live video services
multiview live events
second-screen services
7 / 15 Gwendal Simon Scadoosh

19. Scadoosh
8 / 15 Gwendal Simon Scadoosh

20. Main idea in a nutshell
To not send all representations to all edge servers
9 / 15 Gwendal Simon Scadoosh

21. Formal Scadoosh
e
uij = utility score of representation i of
channel j for edge server e
10 / 15 Gwendal Simon Scadoosh

22. Formal Scadoosh
e
uij = utility score of representation i of
channel j for edge server e

e
 1 if repr i of chn j is delivered to e
xij = 
0 otherwise
10 / 15 Gwendal Simon Scadoosh

23. Formal Scadoosh
e
uij = utility score of representation i of
channel j for edge server e

e
 1 if repr i of chn j is delivered to e
xij = 
0 otherwise
objective :
e e
max uij × xij
e j i
10 / 15 Gwendal Simon Scadoosh

24. Scadoosh Principle
Scadoosh
coordinator
11 / 15 Gwendal Simon Scadoosh

25. Scadoosh Principle
1
edge servers
Scadoosh
report to
coordinator Scadoosh
coordinator about
their activities
during last period
1
11 / 15 Gwendal Simon Scadoosh

26. Scadoosh Principle
2
2
Scadoosh
coordinator
Scadoosh
coordinator
decides utility
scores, computes
delivery forest
overlay and
notifies sources
11 / 15 Gwendal Simon Scadoosh

27. Scadoosh Principle
3
3
sources use the
Scadoosh
coordinator
forest overlays to
deliver the live
representations to
the edge servers
11 / 15 Gwendal Simon Scadoosh

28. What we have done
Formulate an optimization problem
Pascal Frossard (EPFL)
Hervé Kerivin (Clemson then Clermont-Ferrand)
12 / 15 Gwendal Simon Scadoosh

29. What we have done
Formulate an optimization problem
Pascal Frossard (EPFL)
Hervé Kerivin (Clemson then Clermont-Ferrand)
Design optimal and approximate algorithms
Jimmy Leblet (Lyon) and Fen Zhou (Avignon)
Hanan Shpungin (University of Waterloo)
12 / 15 Gwendal Simon Scadoosh

30. What we have done
Formulate an optimization problem
Pascal Frossard (EPFL)
Hervé Kerivin (Clemson then Clermont-Ferrand)
Design optimal and approximate algorithms
Jimmy Leblet (Lyon) and Fen Zhou (Avignon)
Hanan Shpungin (University of Waterloo)
Design and test a system
Catherine Rosenberg (University of Waterloo)
Jiayi Liu (Telecom Bretagne)
12 / 15 Gwendal Simon Scadoosh

31. Results Overview
ratio of received encodings
1 1
0.8 0.8
0.6 mobile 0.6 xdsl
0.4 0.4
0.2 0.2
0 0
120%100% 80% 60% 40% 20% 120%100% 80% 60% 40% 20%
provisioning provisioning
1
0.8 lowest representation
0.6 ftth low representation
0.4 normal representation
0.2
high representation
highest representation
0
120%100% 80% 60% 40% 20%
13 / 15 Gwendal Simon Scadoosh

32. Conclusion
14 / 15 Gwendal Simon Scadoosh

33. Conclusion
Rate-adaptive video streaming :
widely adopted in video services
threatening CDN infrastructure
15 / 15 Gwendal Simon Scadoosh

34. Conclusion
Rate-adaptive video streaming :
widely adopted in video services
threatening CDN infrastructure
Scadoosh is in preliminary stage :
reduce infrastructure by a factor of 5
maintain quality of experience
15 / 15 Gwendal Simon Scadoosh

35. Conclusion
Rate-adaptive video streaming :
widely adopted in video services
threatening CDN infrastructure
Scadoosh is in preliminary stage :
reduce infrastructure by a factor of 5
maintain quality of experience
Stay tuned :
http ://perso.telecom-bretagne.eu/gwendalsimon
15 / 15 Gwendal Simon Scadoosh