End-to-End TCP-Friendly Streaming Protocol and Bit Allocation for Scalable Video Over Wireless Internet
Fan Yang, Qian Zhang, Wenwu Zhu, and Ya-Qin Zhang
IEEE Journal on Selected Areas in Communications, May 2004
Outline Introduction Design issues
Packet loss Delay variation TCP-friendly
WMSTFP Estimation of network conditions End-to-end rate distortion
Simulation results
Video streaming over wireless Internet
Design issues of streaming over wireless Internet The end-to-end packet loss can be caused by
the congestion loss occurred in the wired network the erroneous loss occurred in the wireless
network The variation in end-to-end delay is large.
Packet loss ratio and round-trip time (RTT) is usually used by streaming protocol to adjust sending rate.
The streaming protocol should be friendly to TCP.
System architecture (1/2)
System architecture (2/2) WMSTFP congestion control (sender)
Adjust sending rate based on the feedback information.
WMSTFP network monitor (receiver) Analyze the erroneous loss rate (wireless) and
congestive loss rate (wired). Estimate the end-to-end available network
bandwidth. Network-adaptive ULP channel encoder
Protect different layers of PFGS according to their importance and network status using RS codes.
Loss differentiated R-D-based bit allocation Make the total sending rate adapt to the estimated
network conditions.
Features of WMSTFP Accurate loss differentiation
Detect packet losses caused by the erros in wireless channels using the information acquire at the link-layer.
Forward loss ratio estimation Packets have different loss patterns. (different l
oss burtiness lengths)
Smoothed RTT measurement
Overview of the protocol The process of the protocol:
estimating loss rate (congestive and erroneous); estimating RTT and retransmission time out
(RTO); estimating the available network bandwidth and
adjusting the sending rate.
After slow-start, the sender adjusts its sending rate based on the congestive packet loss ratio, RTT, and RTO.
RTT measurement
RTT and RTO estimation (1/2)
is set to 0.75
avoid the clock synchronization issue
RTT and RTO estimation (2/2)
k is set to 4
roundlast in the RTT estimated theis RTT
RTT of variation theof estimation smoothed theis RTTVAR
RTT. estimatedcurrent theis RTT
'
is set to 0.25
End-to-end packet loss differentiation and measurement Use the link-layer information to differentiate the
wireless erroneous loss and congestive loss.
In the third-generation (3G) wireless communication system, we can deduce a packet loss caused by wireless errors based on the information provided in the radio link control layer (RLC).
We can even get more detailed statistical information such as frame error rate at the radio resource control layer (RRC).
Network-adaptive ULP Applying ULP scheme to different layer to
provide prioritized transmission.
When the network is in good status, more bit budget should be assigned for source coding and fewer bits should be assigned for channel coding.
On the contrary, when network condition is bad, it is necessary to allocate more bits for channel coding, thus fewer bits should be allocated for source coding.
Loss patterns (1/3)
Loss patterns (2/3)
Different loss patterns have different impact on the perceived QoS quality in video streaming.
Loss patterns (3/3)
Comparisons of the 55th reconstructed frame under different schemes.
Assume the loss pattern is randomly distributed.
End-to-end rate distortion
DT: end-to-end distortion
Ds: source distortion (caused by quantization & rate control)
Dc: channel distortion (caused by packet loss)
Rs: source coding rate
Rc: channel coding rate
Problem formulation (1/2) Allocate the available bit rate such that the optima
l Rs and Rc are obtained by minimizing end-to-end distortion under the constraint Rs + Rc ≤ RT. (RT is the estimated network bandwidth)
Problem formulation (2/2)
Pc, layer(i, j) is the probability that the jth packet in the ith layer is lost due to congestion.
Pw, layer(i, j) is the probability that the jth packet in the ith layer is lost due to wireless errors.
not congested & wireless error
Simulation topology
Comparisons of throughput for TCP and WMSTFP connections
Why TCP is better than WMSTFP?
Sending rate comparisons of WMSTFP and TCP
FER = 0.1
Throughput comparisons under different FER
FER = 0.2 FER =
0.3
Why WMSTFP is now better than TFRC?
TFRC (TCP-friendly rate control protocol)
Overall packet loss rate under different FER
TFRC has the higher loss rate.
TCP has the higher throughput?
Performance of loss differentiated R-D-based bit allocation scheme
FULP-T fixed ULP without loss pattern
differentiation over TFRC FULP-W
fixed ULP without loss pattern differentiation over WMSTFP
AULP-W adaptive ULP over WMSTFP
Average PSNR of different schemes under different bit rates
FER = 0.3FER = 0.2
Average PSNR comparisons for foreman using different schemes
FER = 0.3
FER = 0.2
PSNR comparisons for foreman using different schemes
FER = 0.3
FER = 0.2
Comparison of the reconstructed frames under different FERs (1/3)
FER = 0.3
FER = 0.2
AULP-W
Comparison of the reconstructed frames under different FERs (2/3)
FER = 0.3
FER = 0.2
FULP-W
Comparison of the reconstructed frames under different FERs (3/3)
FER = 0.3
FER = 0.2
FULP-T