Packet Loss Recovery and Control for

Packet Loss Recovery and Control for

Voice Transmission over the
Voice Transmission over the
Internet
Internet
Dipl.-Ing. Henning Sanneck
Motivation: impact of packet losses on voice
Approach: combined loss recovery and control
Results:
end-to-end loss recovery
hop-by-hop loss control
combination
Conclusions
October 10, 2000
Packet Loss Recovery and Control for Voice Transmission over the Internet 2
Packet Loss
Packet Loss
in
in
Voice over
Voice over
IP
IP
Networks
Networks
Two major technical evolutions
digital voice communication
packet-switched networks
Main advantages over circuit-switched voice:
statistical multiplexing gain
service integration,
unified packet-switching infrastructure
fundamental tradeoff of best effort network:
statistical multiplexing vs. reliability of transmission
packet loss significant distortions
previous approaches to the packet loss problem:
either end-to-end-only or hop-by-hop-only
Motivation
Packet Loss Recovery and Control for Voice Transmission over the Internet 3
Combined Loss
Combined Loss
Recovery
Recovery
and
and
Control
Control
Approach
best effort
Example:
flow consists of
50% more (+1) and
50% less (-1)
important packets
intra-flow
loss control
loss recovery with
50 100
50
100
loss rate [all packets] (%)
l
o
s
s

r
a
t
e

-
1

p
a
c
k
e
t
s

(
%
)
50 100
50
100
loss rate [all packets] (%)
l
o
s
s

r
a
t
e

+
1

p
a
c
k
e
t
s

(
%
)
utility (%)
overall utility
utility contribution
of the -1 packets
of the +1 packets
50 100
50
100
loss rate [all packets] (%) loss rate [all packets] (%)
50 100
50
100
l
o
s
s

r
a
t
e

-
1

p
a
c
k
e
t
s

(
%
)
l
o
s
s

r
a
t
e

+
1

p
a
c
k
e
t
s

(
%
)
utility (%)
Packet Loss Recovery and Control for Voice Transmission over the Internet 4
Questions
Questions
1 Which packets are more / less important for the user
perception of voice (utility) ?
identification / recovery at the end-to-end level
2 How can the loss distribution (+1, -1) rather than the
absolute amount of loss be controlled within the
network ?
loss control at the hop-by-hop level
3 How can the end-to-end and hop-by-hop mechanisms
be combined ?
Approach
Packet Loss Recovery and Control for Voice Transmission over the Internet 5
Overview
Overview
Approach
emulation
implementation /
network
measurement simulation
discrete event
speech quality model
Hop-by-hop loss control
user-level
model
packet loss model
End-to-end loss recovery
traffic
speech quality /
packet loss
relationship
loss recovery
Combined
and control
simulation
Packet Loss Recovery and Control for Voice Transmission over the Internet 6
End
End
-to-end
-to-end
identification
identification
/
/
recovery
recovery
Results
a
m
p
l
i
t
u
d
e
time
[ms]
0 50 100 150 200 250 300
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0 300
importance
frame-based compression
packetization
loss
compression recoverable
loss pattern
sample alternating
{+1,-1,...}
frame bursty
{+1,+1,-1,-1}
Speech Property Based (SPB)
identification & protection /
internal decoder loss concealment
Packet Loss Recovery and Control for Voice Transmission over the Internet 7
Hop
Hop
-
-
by
by
-
-
hop loss control
hop loss control
sender marks packets with
recoverable loss pattern
Results
Differential Random Early
Detection (DiffRED) enforces
the pattern
well-defined drop probability
relationship between +1, -1:
still best effort
example:
alternating pattern {+1, -1,...}
problem: simple packet loss
models cannot capture
algorithm behaviour
new model
1 2 3 4 5 6 7 8 9 10
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
burst length k
b
u
r
s
t

l
o
s
s

l
e
n
g
t
h

r
a
t
e

g
k
,

e
s
t
i
m
a
t
e
d

b
u
r
s
t

l
o
s
s

l
e
n
g
t
h

run-length model (m=2)
Drop Tail
DiffRED
Drop Tail (estimated)
DiffRED (estimated)
1
1 10
Drop Tail
Drop Tail
(model)
DiffRED
(model)
DiffRED
burst length
PDF
1
of loss burst lengths
10
-6
1
ideal
1
Probability Density Function
Packet Loss Recovery and Control for Voice Transmission over the Internet 8
Combined loss recovery and control
Combined loss recovery and control
Results
hop-by-hop-
only (p
+1
=10
-3
p
0
)
end-to-end-
only
combined
(best effort:
p
-1
= 2 p
0
- p
+1
)
p
+1
p
0
p
-1
p
+1
p
0
0 0 0 0 0 0 NO MARK
p
0
SPB DIFFMARK
ALT DIFFMARK
0
+1
0 0 +1 +1
+1 +1 +1
-1 -1
-1 -1 -1
-1
FULL MARK +1 +1 +1 +1 +1 +1
SPB & DiffRED:
SPB-FEC
end-to-end
algorithm
hop-by-hop
algorithm
1
Speech Property Based loss protection
2
Forward Error Correction
1

2

Packet Loss Recovery and Control for Voice Transmission over the Internet 9
0.05 0.1 0.15 0.2 0.25
-5
-4.5
-4
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
Utility (
Utility (
user perception
user perception
)
)
( (objective speech quality measurement objective speech quality measurement) )
excellent
good
fair
0
0.05 0.25
SPB-FEC
SPB-DIFFMARK
ALT-DIFFMARK
NO MARK
FULL MARK
drop probability p
0
E
M
B
S
D

P
e
r
c
e
p
t
u
a
l

D
i
s
t
o
r
t
i
o
n
Results
5
Packet Loss Recovery and Control for Voice Transmission over the Internet 10
Conclusions
Conclusions
Combined mechanisms:
performance improvement for best effort networks
end-to-end:
no redundancy / feedback / adaptivity,
knowledge about the codec, processing needed
hop-by-hop:
no charging architecture (intra-flow QoS),
deployment in routers necessary
End-to-end mechanisms:
selective (speech property-based) loss protection
comes close to protection of the entire flow
Summary
Packet Loss Recovery and Control for Voice Transmission over the Internet 11
Conclusions
Conclusions
Hop-by-hop mechanisms:
trading between losses with simple mechanisms in a
best effort network
Measurement of loss process / loss impact:
novel packet loss model
application of perceptual metrics at the user level
Applicability validated in real Internet environment
Future: assessment of large scale deployment costs
Summary