Post on 22-Feb-2016
description
transcript
1NUS.SOC.CS5248OOI WEI TSANG
Playout Buffer Management
NUS.SOC.CS5248OOI WEI TSANG 2
You are Here
Network
Encoder
SenderMiddlebox
Receiver
Decoder
NUS.SOC.CS5248OOI WEI TSANG 3
How to recv and play?open socketwhile not done
if socket is readable read packet from socketremove RTP headerdecodeplay back
NUS.SOC.CS5248OOI WEI TSANG 4
What’s Wrong?packet orderingpacket lossnext packet arrive in-time?
Especially bad for audio applications
5NUS.SOC.CS5248OOI WEI TSANG
Detour: A Brief Intro to Audio Conferencing
NUS.SOC.CS5248OOI WEI TSANG 6
Audio CompressionNormally uncompressed
Telephone quality: 8-bit audio 8Khz
20-30ms per packet
NUS.SOC.CS5248OOI WEI TSANG 7
Common TechniqueSilence Detection
No need to send if there is no sound at the input
Talkspurt Sequence of consecutive audio
packets (in between silence)
NUS.SOC.CS5248OOI WEI TSANG 8
Detecting Talkspurt in RTPMarker bit
Not reliable as packet with marker bit could be lost
Deduce from timestamp and sequence number
9NUS.SOC.CS5248OOI WEI TSANG
Delay Jitter
NUS.SOC.CS5248OOI WEI TSANG 10
What causes Jitter?Network delay =
Propagation Delay (fixed) +Queueing Delay (variable)
Delay jitter is caused by variable queueing delay
NUS.SOC.CS5248OOI WEI TSANG 11
Delay Jitter
Time
TransitTime
small jitter
large jitter
NUS.SOC.CS5248OOI WEI TSANG 12
Spike
Time
TransitTime
13NUS.SOC.CS5248OOI WEI TSANG
Buffer: The Jitter Absorber
NUS.SOC.CS5248OOI WEI TSANG 14
If Zero Jitter
Time
NUS.SOC.CS5248OOI WEI TSANG 15
With Jitter
TimeLOSS
NUS.SOC.CS5248OOI WEI TSANG 16
With Jitter + Buffer
Time
Buffer Size
Playout Delay
17NUS.SOC.CS5248OOI WEI TSANG
Q: How to set playout delay?
NUS.SOC.CS5248OOI WEI TSANG 18
Adapting Playout DelayWhen jitter is low, reduce delayWhen jitter is high, increase delay
NUS.SOC.CS5248OOI WEI TSANG 19
Fundamental Trade-off
Latency vs Packet Loss
NUS.SOC.CS5248OOI WEI TSANG 20
Playout DelayOnce you set the playout delay,
cannot change!
NOT true: can change at beginning of talkspurt
NUS.SOC.CS5248OOI WEI TSANG 21
Adapting Playout Delay
SEND
RECV
PLAY
22NUS.SOC.CS5248OOI WEI TSANG
Adaptation Algorithm
NUS.SOC.CS5248OOI WEI TSANG 23
Variables and Notations
Tsend(i)
Tplay(i)
Tbuffer(i)
Tarrive(i)
Tdelay(i)
Tnet(i)
NUS.SOC.CS5248OOI WEI TSANG 24
First Packet in the TalkspurtTdelay(i) = Enet(i) + 4 Vnet(i)Tplay(i) = Tsend(i) + Tdelay(i)
Tsend(i)
Tplay(i)
Tbuffer(i)
Tarrive(i)
Tdelay(i)
Tnet(i)
NUS.SOC.CS5248OOI WEI TSANG 25
Subsequent PacketsTplay(j) = Tplay(i) + Tsend(j) – Tsend(i)
Tsend(i)
Tplay(i)
Tbuffer(i)
Tarrive(i)
Tdelay(i)
Tnet(i)
26NUS.SOC.CS5248OOI WEI TSANG
How to estimate Enet(i)
NUS.SOC.CS5248OOI WEI TSANG 27
Algorithm 1 (Jacobson’s)Enet(i) = aEnet(i-1) + (1-a)Tnet(i)
NUS.SOC.CS5248OOI WEI TSANG 28
Algorithm 2 if Tnet(i) > Enet(i)
Enet(i) = bEnet(i-1) + (1-b)Tnet(i)
else
Enet(i) = aEnet(i-1) + (1-a)Tnet(i)
NUS.SOC.CS5248OOI WEI TSANG 29
Algorithm 3Enet(i) = min { Tnet(j) }
(over all packets in previous talkspurt)
NUS.SOC.CS5248OOI WEI TSANG 30
Algorithm 4 (Ramjee’s)Ramjee’s Proposal
Observation: Algorithm 1-3 take too long to react to spike.
NUS.SOC.CS5248OOI WEI TSANG 31
Spike
NUS.SOC.CS5248OOI WEI TSANG 32
Ramjee’s IdeaWorks in 2 modes
SPIKE NORMAL
NUS.SOC.CS5248OOI WEI TSANG 33
Three QuestionsWhen to switch from normal to
spike mode?
When to switch from spike back to normal?
How to estimate during spike mode?
NUS.SOC.CS5248OOI WEI TSANG 34
Normal to Spike if difference in delays is large
NUS.SOC.CS5248OOI WEI TSANG 35
Normal to Spike if difference in delays is large
difference in delays:Tnet(i) – Tnet(i-1)
large:800 + 2Vnet(i)
NUS.SOC.CS5248OOI WEI TSANG 36
Another View of Spike (ZOOM)
Tnet(i)
Tarrive(i)
NUS.SOC.CS5248OOI WEI TSANG 37
Spike to Normal slope = slope/2 +
|2Tnet(i) – Tnet(i-1) – Tnet(i-2)|/8
if slope < 64 switch to normal
NUS.SOC.CS5248OOI WEI TSANG 38
Estimation in Spike ModeEnet(i) = Enet(i-1) + Tnet(i) – Tnet(i-1)
NUS.SOC.CS5248OOI WEI TSANG 39
First Packet in the TalkspurtTdelay(i) = Enet(i) + 4 Vnet(i)Tplay(i) = Tsend(i) + Tdelay(i)
NUS.SOC.CS5248OOI WEI TSANG 40
First Packet in the TalkspurtTdelay(i) = Enet(i) + 4 Vnet(i)Tplay(i) = Tsend(i) + Tdelay(i)
Tsend(i)
Tprop(i)
Tplay(i)
Tbuffer(i)
Tarrive(i)
Tq(i)
Tdelay(i)
Tnet(i)
41NUS.SOC.CS5248OOI WEI TSANG
How to estimate Vnet(i)
NUS.SOC.CS5248OOI WEI TSANG 42
Variation of DelayVnet(i) =
aVnet(i-1) + (1-a)|Enet(i) – Tnet(i)|
43NUS.SOC.CS5248OOI WEI TSANG
Comparisons of 4 Algorithms
NUS.SOC.CS5248OOI WEI TSANG 44
Playout Delay vs Loss Rate
Tdelay(i)
Loss Rate
45NUS.SOC.CS5248OOI WEI TSANG
Algorithm 5 (Moon’s)
NUS.SOC.CS5248OOI WEI TSANG 46
Problems of Existing Algorithms Jacobson’s react too slowlyRamjee’s follow the delay too
closely
NUS.SOC.CS5248OOI WEI TSANG 47
Moon’s IdeaCollect statistics on packets that
have arrived
Find t such that q% of last w packets has Tnet(i) < t
Tdelay(i) = t
NUS.SOC.CS5248OOI WEI TSANG 48
Example (w =50, q = 90%)
num of packets
delay 1 2 3 4 5 6 7 8 9 10 11 12
NUS.SOC.CS5248OOI WEI TSANG 49
Spike ModeTdelay(i) = Enet(1)
SPIKE NORMAL
NUS.SOC.CS5248OOI WEI TSANG 50
Spike Mode
NUS.SOC.CS5248OOI WEI TSANG 51
Switching ModeNormal to Spike
if (Tnet(i) > k*Tdelay(i))
Spike to Normal
if (Tnet(i) < k’ * OLDTdelay(i))
NUS.SOC.CS5248OOI WEI TSANG 52
Loss Rate vs. DelayTdelay(i)
Loss Rate
53NUS.SOC.CS5248OOI WEI TSANG
Theoretical Lower Bound
NUS.SOC.CS1102OOI WEI TSANG
54
Dynamic Programming
NUS.SOC.CS1102OOI WEI TSANG
55
Fibonacci Numbers
0 1 …
0 1 2
…
i-2 i-1 i
otherwiseii
FFF
ii
i 10
10
21
+
NUS.SOC.CS1102OOI WEI TSANG
56
Fibonacci Numbersfib(n)
x[0] = 0x[1] = 1for (i = 2; i < n; i++)
x[i] = x[i-1] + x[i-2]return x[i]
NUS.SOC.CS1102OOI WEI TSANG
57
Binomial Coefficient
nor 0 if 1
111
kkn
kn
kn 1
1 1
1 1
1 1
1
1
k
n
NUS.SOC.CS1102OOI WEI TSANG
58
Change-Making Problem [Weiss] 7.6
For a currency with coin C1, C2, .. Cn (cents), what is the min number of coins needed to make K cents of change?
NUS.SOC.CS1102OOI WEI TSANG
59
Example C = {1, 5, 10, 20, 50} K = 76 cents Give 4 coins 50 + 20 + 5 + 1 = 76
NUS.SOC.CS1102OOI WEI TSANG
60
Formulation To make a change of K cents, either
make a change of (K-50) cents, or make a change of (K-20) cents, or make a change of (K-10) cents, or make a change of (K-5) cents, or make a change of (K-1) cents
Number of coins for K = 1 + minimum of all the above choices
NUS.SOC.CS1102OOI WEI TSANG
61
Dynamic Programming)}({min1)( ii
CKcoinUsedKcoinUsed
… …
min +1
KK-5K-10K-20
62NUS.SOC.CS5248OOI WEI TSANG
Theoretical Lower Bound
NUS.SOC.CS5248OOI WEI TSANG 63
Goal Input:
A packet trace A number of packet losses
What is the minimum average playout delay Tdelay(.) ?
NUS.SOC.CS5248OOI WEI TSANG 64
Trace
kM talkspurts
1,k 2,k 3,k j,k nk,k
NUS.SOC.CS5248OOI WEI TSANG 65
New NotationsM:
Number of TalkspurtNpacket(k) or nk
Number of packets in talkspurt kNtotal
Total number of packets
NUS.SOC.CS5248OOI WEI TSANG 66
Define..d(k, i)
minimum average playout delay for choosing i packets to be played out from k-th talkspurt
k.. M..
1,k 2,k 3,k j,k nk,k
NUS.SOC.CS5248OOI WEI TSANG 67
Define..D(k, i)
minimum average playout delay for choosing i packets to be played out from k-th to M-th talkspurt
k.. M..
NUS.SOC.CS5248OOI WEI TSANG 68
Dynamic Programming
1
2
3
M
0 1 2 N
NUS.SOC.CS5248OOI WEI TSANG 69
Formulation for Base Caseif i = 0 then
D(k,i) = 0if k = M
if i ≤ Npacket(M) thenD(k,i) = d(k,i)
elseD(k,i) = ∞
NUS.SOC.CS5248OOI WEI TSANG 70
Formulation for Recursive Case
k.. M..
i
jkjdjikDjiikDij
),(),1()(min),(0
NUS.SOC.CS5248OOI WEI TSANG 71
Loss Rate vs. DelayTdelay(i)
Loss Rate1%
72NUS.SOC.CS5248OOI WEI TSANG
Video Playout
NUS.SOC.CS5248OOI WEI TSANG 73
Two Methods I-Policy
Display at fixed playout delay Drop late frames
E-Policy Display late frame at next period Playout delay increase
NUS.SOC.CS5248OOI WEI TSANG 74
I-Policya b c d e f g h
1 2 3 4 5 6 7 8 9 10 a c e f h
NUS.SOC.CS5248OOI WEI TSANG 75
E-Policya b c d e f g h
1 2 3 4 5 6 7 8 9 10 11 a b d f gc e h
NUS.SOC.CS5248OOI WEI TSANG 76
I-Policya b c d e f g h
1 2 3 4 5 6 7 8 9 10 a b e f hg
NUS.SOC.CS5248OOI WEI TSANG 77
E-Policya b c d e f g h
1 2 3 4 5 6 7 8 9 10 a b c d fe
NUS.SOC.CS5248OOI WEI TSANG 78
Pro and Cons I-Policy: Fixed latencyE-Policy: No Loss Frame or Gap
NUS.SOC.CS5248OOI WEI TSANG 79
Ideaif queue length > L for T seconds
drop incoming frames
long queue -> small Tshort queue -> large T
NUS.SOC.CS5248OOI WEI TSANG 80
Queue Monitoring
3
2
3
3
3
4
2
5
2
6
2
7
2
8
0 0 0 0 0 0 0
queue length
threshold
waiting time
NUS.SOC.CS5248OOI WEI TSANG 81
Queue Monitoring
3
2
3
3
3
4
2
5
2
6
2
7
2
8
0 0 0 0 0 0 0
queue length
threshold
waiting time
NUS.SOC.CS5248OOI WEI TSANG 82
Queue Monitoring
3
2
3
3
3
4
2
5
2
6
2
7
2
8
1 0 0 0 0 0 0
queue length
threshold
waiting time
NUS.SOC.CS5248OOI WEI TSANG 83
Queue Monitoring
3
2
3
3
3
4
2
5
2
6
2
7
2
8
2 1 0 0 0 0 0
queue length
threshold
waiting time
NUS.SOC.CS5248OOI WEI TSANG 84
Queue Monitoring
3
2
3
3
3
4
2
5
2
6
2
7
2
8
3 2 1 1 0 0 0
queue length
threshold
waiting time
NUS.SOC.CS5248OOI WEI TSANG 85
SummaryPreliminary work by Stone and
Jeffay on Video Playout
Ramjee’s and Moon’s adaptive audio playout algorithm
86NUS.SOC.CS5248OOI WEI TSANG
Possible Survey and Project Topic
NUS.SOC.CS5248OOI WEI TSANG 87
Adapting PlaybackBesides adapting buffering time,
we can adapt playback time
Estimating and eliminating clock drift