Title: Optimized Transcoding Rate Selection and Packet Scheduling for Transmitting Multiple Video Streams O
1Optimized Transcoding Rate Selection and Packet
Scheduling for Transmitting Multiple Video
Streams Over a Shared Channel
Mark Kalman, Peter van Beek and Bernd Girod
2Overview
- Application DTV over wireless LAN
- Finding Optimal Transcoding Rates
- Effect of Transcoder Control Lag
- Source Pruning and Packet Scheduling
- Distortion Model
- Delay Model
- Experimental results
3Distribution of DTV over WLAN
diagram taken from van Beek, 2004
4Video Streaming Over Shared Channel
Transcoder
Decoder
0
0
Transcoder
Decoder
1
1
Receiver
Transcoder
Decoder
(Multi-Channel)
2
2
Transcoder
Decoder
3
3
Controller
5Video Streaming Over Shared Channel
Transcoder
Decoder
0
0
Transcoder
Decoder
1
1
Receiver
Transcoder
Decoder
(Multi-Channel)
2
2
Transcoder
Decoder
3
3
6Server Divides Channel Time Among Links
7Link Capacities Vary Over Time
8Transcoders Control Source Rate for Each Link
Transcoder
Decoder
0
0
Transcoder
Decoder
1
1
Receiver
Transcoder
Decoder
(Multi-Channel)
2
2
Transcoder
Decoder
3
3
9R-D for H.264 Transcodings of mobile.cif
10Choosing Optimal Transcoding Rates
- Number of streams
- Fraction of channel time devoted to stream m
- Capacity of link m
- Weights the importance of the distortion of
stream m
- Rate-Distortion function for stream ms source
- Transcoded rate of source m
11Closed-Form Solution using Lagrange Multipliers
(1)
Zhu and Setton 2004
Stuhlmueller, Faerber, Link, and Girod 2000
12Closed-Form Solution using Lagrange Multipliers
(2)
13Transcoders Support Discrete Rates
Iterate to find discrete solution
14Simulated Link Capacities
Link 1
Link 0
Link 2
Link 3
15Channel Time Allocations, Sm
Link 0
Link 1
Link 3
Link 2
16Transcoded Rate
Link 1
Link 0
Link 3
Link 2
17Link 2 Capacity, Sm, Rm - Group Shot
18Backlog of Transmission Queues
Transcoder
Decoder
0
0
Transcoder
Decoder
1
1
Receiver
Transcoder
Decoder
(Multi-Channel)
2
2
Transcoder
Decoder
3
3
Controller
19Backlog of Transmission Queues
20Effect of Transcoder Control Lag (1)
Capacity High
Transcoder
3
Controller
21Effect of Transcoder Control Lag (2)
Capacity Drops!
Transcoder
3
Controller
22Effect of Transcoder Control Lag (3)
23Transcoded Rate
Link 1
Link 0
Link 3
Link 2
24No Lag Y-PSNR 35 dB
66 ms Lag Y-PSNR 27 dB
25Backlog of Transmission Queues
Transcoder
Decoder
0
0
Transcoder
Decoder
1
1
Receiver
Transcoder
Decoder
(Multi-Channel)
2
2
Transcoder
Decoder
3
3
Controller
26Pruning and Scheduling
B
I
B
B
B
P
B
B
B
P
B
B
B
P
Goal Choose packets for dropping that will
approximately maximize expected PSNR
B
P
B
B
P
B
B
I
B
P
Controller
27PSNR Vs. Frame Omission Pattern
B
B
B
B
P
I
P
B
B
B
B
P
P
B
B
28Calculate Probability Of Frame Loss
- Frame arrival deadline
- Number of successful packet transmissions
needed to complete a frame
- Time until timeout
- R.V. modeling the number of timeouts until n
successes if failure (time out) probability is e
(Neg. Binomial Distributed)
- Fraction of channel time devoted to stream m
- R.V. modeling time to complete packet
transmission (Exp. distributed)
- Bernoulli probability that packet transmission
times out
29Calculating Estimated PSNR
30Scheduling Algorithm
31Testing Trace Collection Accuracy
802.11b
media packet
Wired Ethernet
Trace client
Trace server
Synch packets
If wireless channel is in bad state, sends on
wired link are delayed. We dont know if delays
are due to wireless channel or due to timing
inaccuracy
32Flooding with Orinoco 11mbps cards
- Stations separated by 15 meters
- Media packets transmitted continually
- Delay spikes of 15 20 ms seen maybe on a 10-15
second basis
33Detail of Flooding Trace
34Relative Frequencies of Xm,i
35Relative Frequencies of Delays
Gamma approximation has min KS distance DKS
0.0891
36Pruning Performance
37Scheduling And Transcoder Lag
38Scheduling 66 ms lag Y-PSNR 34 dB
Round-Robin 66 ms lag Y-PSNR 27 dB
39Conclusion
- Closed-Form solution for optimal rates
- Transcoder control lag leads to distortion
- Optimized pruning and packet scheduling
- Simple additive distortion model (No DAG)
- Sum of Packet Delays Modeled as G or Gaussian
- Pruning Performance approaches R-D Optimal
- Scheduling yields over 6 dB gain in simulations