1. Optimising Streaming Systems with SDN/P4/NetFPGA
T. Khoa Phan
University College London
NetFPGA Summit, Cambrige, April 2017.

2. Facebook Livestream System
Origin server
Edge cache
Live stream server
INTERNET
Edge cache
Origin server
Looking at this pic, the bottle neck now is between the edge cache and users because the edge cache has to send every stream for every user. Given almost 2 billions active FB users, this scalability is still not enough. The question is can we design a network protocol for the edge cache to serve users in a more efficient way. To do that, let me zoom in the connection between the edge cache and users.
Facebook livestream system – Content Delivery Network [1]
Each edge cache can serve up to 200,000 users simultaneously
[1]

3. Unicast vs. Multicast Network
Edge cache
Today’s unicast network
Edge cache
IP Multicast
Edge cache
Xcast Treemap
First standardized in 1986.
One stream for all receivers.
Need multicast routers everywhere (or complex tunneling techniques).
Can work for any streaming systems.
Source code: github.com/ptkhoa1984/Xcast6_Treemap

4. Xcast treemap part (optional)
What is Xcast Treemap? S A B C D E
Breath-first tree traversal
A B C D E
List of IP addresses 2
Treemap
Sender S creates packets:
A B C D E
Src: S, Dest: A
Payload 2
Unicast part
Xcast treemap part (optional)
Today router only understands unicast part.
Xcast router lookups and forwards for each IP in the list.

5. How Xcast Treemap works?
Unicast routing table at R3
Dest
Next hop C D E
Full Xcast packet header created by S: D A B
A | A B C D E
D | D
B | B S C
Today routers don’t see the list of destination. To sum up: for today Internet, if users install Xcast Treemap in the computers, it will send data like the overlay tree defined by the sender. R1 R2 R3
C | C D E
E | E
C | C D E
Dest
Next hop A
B, C, D, E R2 E
Today router
Fig. 1: only today routers
Unicast routing table at R1

6. How Xcast Treemap works?
Unicast routing table at R3
Dest
Next hop C D E
Full Xcast packet header created by S: B D A
A | A B C D E
D | D
B | B
C | C S C R1 R2 R3
C doesn’t need to forward ptk to D and E. It has been done automatically by Xcast router and network layer multicast is much better than P2P
My messages:
no matter if there’re Xcast router on the network or not. If there are, no matter where they are, Xcast guarantee that all hosts will receive the pkt correctly.
Xcast Treemap is a stateless protocol: no states at routers and no state at end-hosts, everything is in the packet header.
C | C D E
E | E
C | C D E
Dest
Next hop A
B, C, D, E R2 E
Xcast router
Today router
Unicast routing table at R1
Fig. 2: mixture of Xcast and today routers

7. Facebook Livestream System
current system
with Xcast Treemap
Edge cache server and users

8. Xcast Treemap with SDN S A B C S A B C Shortest path routing
SDN controller S B A C R2 R1 S A X3 B C
Overlay tree by SDN controller
Traffic engineering for Xcast traffic

9. Future work with P4/NetFPGA
Current approach for Xcast router:
Plugin additional network cards for a computer
Install Xcast router software on FreeBSD
Future work for Xcast router:
P4/C#/NetFPGA Xcast router
Other possible platforms?

10. Deployment Scenarios With Xcast NetFPGA gateway
Only with Xcast end-hosts
With Xcast backbone routers

11. Thank you!!! Email: t.phan@ucl.ac.uk

12. Xcast Treemap Evaluation
In P2P, packet loss happens when end-hosts leave (e.g. A leaves and B replaces A in the tree – hosts need to update fwd table in their memory).
Without Xcast router: sender S sends pkt with new treemap header, there’s no fwd table at hosts.
With all Xcast router: pkts are replicated at routers, no pkt loss when hosts join/leave. S A
P2P system B C
P2P system
Xcast: without Xcast router
All routers are Xcast routers

13. Xcast Treemap with SDN S A B C S A B C Shortest path routing
SDN controller S B A C R2 R1 S A X3 B C
Overlay tree by SDN controller
Traffic engineering for Xcast traffic