Multimedia Transmission over IP

1
Multimedia Transmission over IP Trade-off between
channel loss and queuing loss
Seyed Bahram Zahir-Azami Carleton university,
Ottawa September 12, 2000
2
General Diagram

• Voice over IP (VoIP) attracts much attention
  because of the desire of having a single service
  provider for data, voice and multimedia.

IAD
Network Service Provider
ATM or Frame Relay Transmission
DSLAM
IAD
3
Channel Coding Considerations
DSLAM
ISP
user 1
1 Mbps links
user 2
IAD 1
user N
.
45 Mbps link
to Internet
user 1
user 2
IAD M
user N

• There are two sources for loss congestion and
  channel impairment.
• Channel coding (FEC) increases robustness
  vis-à-vis the channel impairments (error, loss).

4
Packetization-1
IAD
DSLAM
user 1
queue
user 2
45 Mbps link
user N
decoder
encoder
1 Mbps link

• Only speech (G.729A) sources.
• N sources are packed in a frame slot of 10 msec.

5
Packetization-2

• ATM cell transmission ( 424 bits 53 bytes).
• On the 1Mbps link, 25 ATM cells per time slot.
• Delay being the most important restriction
• Retransmission (ARQ) is not allowed.
• Channel coding should be applied on a slot per
  slot basis (a block of information has maximum
  the same length as a frame of 10 msec).

6
Source Generation-1

• Each source is either in ON or in OFF mode.
• When active, either in silence or in speech
  mode.

210 sec.
a
t
10 sec.
t
b
7
Source Generation-2

• Two state-diagrams

PC01
PS01
PC11
PS11
PS00
PC00
OFF
ON
silence
speech
PC10
PS10

• Practical parameter values
• average communication time (CL) 210 s
• communication ratio (CR) 25
• average utterance length (SL) 10 s
• speech ratio (SR) 40

8
Channel Model-1
PG01
PG00
PG11
good
bad
PG10

• Gilbert model represents the variable nature of
  the wireless channel with its different states.
• A simple Gilbert model of two states is
  characterized by two parameters
• - average burst rate (BR)
• - average burst length (BL)

9
Channel Model-2

• Example
• BL 5 ATM cells (? 2 msec.)
• BR 0.01 (1)

10
Channel Model-3

• State transition only at the cell boundaries.
• Good state corresponds to a perfect channel.
• A time slot for the 1 Mbps link 25424 bits.
• Considering loss bursts (in the order of 1 msec),
  a channel coding of all the 25 ATM cells of one
  time slot should be designed for the recovery of
  the losses.

11
Channel Coding

• Reed-Solomon code RS ( )
• m 10 bits/symbol
• n 1023 symbols
• Rc k/n
• dmin n-k1
• Example
• Rc 0.8 ? k 818
• dmin 206 ? 5 cells can be recovered.
• 20 cells out of 25 are actual information packets
  and the other 5 are parity.

424 bits 1 cell
1 2 3 .. 39
40 41
42 43 44 .. 80
81 82
83 84 85 .. 120 121
122
25 cells
944 945 946 .. 982 983
984
985 986 987 ... 1023
14 bits
12
Loss Rate vs. Channel Coding Rate-1
N 160 BL 1 cell BR 10
optimum
channel loss
queuing loss
Trade-off in packet switching system, decreasing
channel coding rate results in less channel
losses but increases the bit rate and loss rates
due to buffer overflow.
13
Loss Rate vs. Channel Coding Rate-2
N 160 BL 2.5 cells BR 10
optimum
channel loss
queuing loss
For a channel with longer loss bursts (BL2.5,
i.e., 1 msec), the minimum is not as good as the
previous case (BL1, i.e., 400 ?sec).
14
Loss Rate vs. Channel Coding Rate-3
N 180 BL 1 cell BR 10
optimum
queuing loss
channel loss
For larger number of users (N 180), the minimum
is not as good as the first case (N 160).
15
Loss Rate vs. Channel Coding Rate-4
N 180 BL 2.5 cells BR 10
optimum
channel loss
queuing loss
Here, the minimum is again achieved with Rc 0.8
and the loss is around 7 (compared to 10 if no
channel coding is applied). However, the nature
of loss is different.
16
Buffer Management Policies-1

• Subjective tests show that the vocoder is robust
  against loss, if the loss rate is not too high
  and the losses are not too correlated.

MOS
0 1 2 3 4 5
Loss rate ()
0 10 20 30 40 50

• For the queuing loss, a policy can be adopted to
  minimize the correlation of queuing losses in IAD
  level.

17
Buffer Management Policies-2
dropping decision

• Simple
• Random
• Privileged

select the less privileged source based on
G.729A error concealment and user priorities
t-L-1
t-L2
t
t-1
..
user 1
2
3
if more than one packet of this user are in the
queue drop the oldest one
N
18
Buffer Management Policies-3

• In all the policies, the total number of losses
  is the same and only their distribution is
  different.
• There are different ways to analyze the
  performances of these policies
• sample loss patterns
• instantaneous pmf
• autocorrelation of the losses
• consecutive loss pmf

19
Sample Loss Patterns
ON/OFF random quasi periodic
Simple random Privileged
20
Instantaneous pmf
scattered Gaussian form Dirac
Simple random Privileged
21
Autocorrelation Function of the Loss
N 180 BL 2.5 cells BR 10
Queuing policy results in a large drop in the
loss autocorrelation function for small lags
which means that the losses are sparse in time.
22
Consecutive Losses (pmf)
N 180 BL 2.5 cells BR 10
In the same three cases, the probability of
having consecutive losses is the less for the
privileged policy.
23
Conclusions

• Channel coding improves the performance of the
  transmission system, while there is a tradeoff in
  choosing the optimum coding rate.
• The noisier the channel, the shorter the bursts
  or the less sources are active, the more
  important will be the effect of channel coding.
• An appropriate buffer management makes the losses
  more sparse in time.
• The buffer management renders a random and
  undetermined loss pattern to a quasi periodic one
  which is important in future vocoder standards.