1

CARROM: Client Association, Rate and Redundancy

Optimization for Wi-Fi Multicast

Vishnu Navda, Ram Ramjee, Lenin Ravindranath (MSR India)Ranveer Chandra, Jitu Padhye (MSR Redmond)

2

Why Multicast?

• Today– Live video/audio streaming

• Potential– Windows/Virus Update– IPTV– Application Level Multicast leveraging IP Multicast

3

Wi-Fi Multicast Issues

• Data Sent only using Basic Rates (1/2 Mbps)– Resource under utilization

• Rate anomaly– Low rate client => poor

throughput for everyone• No back-off after collision

– Unicast unfairness• No Retransmissions

– High Losses

Data Rate: 1 Mbps

Access Point

Multicast Group

Clients

4

CARROM ArchitectureInter-AP

optimizer

Intra-AP optimizer

Intra-AP optimizer

Intra-AP optimizer

• Optimal Association• Rate Adaptation• Pseudo-Broadcast• Adaptive FEC

5

Problem - Only basic rates

• Multicast sent at only basic rates (1/2Mbps)

1 Mbps

1 Mbps

11 Mbps

5.5 Mbps

11 Mbps

1 Mbps

Solution• Rate adaptation

5.5 Mbps

5.5 Mbps

5.5 Mbps

AP1

C1

C2C3

6

Problem - Rate Anomaly

11 Mbps

1 Mbps

1 Mbps

1 Mbps

1 Mbps

1 Mbps1 Mbps

11Mbps

Solution •Control Association

C1

C2

C3

AP1 AP2Channel X Channel Y

7

Solution - Optimal Association

• Inter-AP optimizer– Collects feedback from clients– Determines optimal association– Inform clients about association– Inform APs about rate

AP1AP2

Inter-AP optimizer

C1C2 C3

8

Optimal Association

• Rate Matrix• Metrics– Minimize sum of multicast air time across APs– Maximize average group throughput

• Optimized Association• C1 - AP1• C2 - AP2• C3 - AP2

AP1 AP2

C1

C2

C3

9

Optimal Association

• NP Hard problem• Hierarchical Association Algorithm– Min Cut to Partition topology– LP solver– Brute force with N best APs when partition size is small

• Greedy Algorithm– Online– Associate clients/groups to that AP which minimizes

impact to existing flows

10

Problem - Unicast unfairness

• When unicast and multicast co-exists– Unicast backs-off– Multicast doesn’t– Degrades unicast throughput

Solution •Pseudo-Broadcast approach

Ref: MSWIM06

11

Solution - Pseudo-Broadcast

• AP unicast to one client• Other clients listen in promiscuous mode

• Unicast fairness due to back-off

Pseudo-broadcastListening in

promiscuous mode

12

Losses are uncorrelated

• Losses are uncorrelated across clients

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Correlation

Time in 30 sec

Co

rrel

atio

n C

oef

fici

ent

Solution•Adaptive FEC

Correlation between 2 users

13

Adaptive FEC

• Intra-AP optimizer– Pseudo-Broadcast• Packet level feedback• Weighted round-robin among clients• Adapt FEC

14

Summary

• Optimal Association– Increase Multicast Throughput

• Pseudo-Broadcast– Unicast Fairness

• Multicast Rate Adaptation– Increase Multicast Throughput

• Adaptive FEC– Increase Reliability

15

Results – Optimal Association

Optimal Association helps significantly reduce Wireless Utilization

1 2 3 40

2

4

6

8

10

12

14

16

18

N best APs in brute force algorithm

Air

time

to t

rans

mit

1M

b (s

ec)

Number of Clients: 20Number of APs: 10Number of Groups: 10

16

Implementation Issues

• No modification– Optimal Association

• AP modification– Rate Adaptation– Adaptive FEC– Pseudo broadcast

• Missing Windows APIs– Associate to a AP with given BSSID– Listen in promiscuous mode