Resource Assignment for Integrated Services in Wireless ...
Transcript of Resource Assignment for Integrated Services in Wireless ...
Resource Assignment for IntegratedServices in Wireless Multimedia Networks
Victor [email protected]
http://research.microsoft.com/~bahl
Microsoft ResearchRedmond, Washington
August 8, 1997
August 8, 1997 Victor Bahl
3UHVHQWDWLRQ�2XWOLQH
1 Introduction and Problem Description
2 Bandwidth Allocation, Reservation, and Utilization→ Connection level Intra-Frame Statistical Multiplexing
3 Bandwidth Partitioning→ Priority Sharing with Restrictions (PSR)
→ The Smart Allocate Algorithm
4 Conclusions
Current Research Activities
End NodeOptimizations
System LevelOptimizations
August 8, 1997 Victor Bahl
,QWURGXFWLRQ�DQG�3UREOHP�'HVFULSWLRQ
l Different broadband sources (traffic classes) require differentamounts of bandwidth and have different transmission priorities
• Bandwidth: Video > Voice > Data
• Priorities: Voice > Video > Data
l Challenge for the network designer: Develop techniques that can assign bandwidth to the different traffic classes
in manner such that• a guarantee on the level of service can be provided to a portion of the
time-critical traffic classes, and
• the needs of the system (i.e. maximize the # of supported connections)are balanced against the needs of the applications
(Bandwidth prediction and reservation for bursty sources such as VBR video isan unsolved problem, so how do we allocate bandwidth for such sources)
August 8, 1997 Victor Bahl
3RVVLEOH�$SSURDFKHV
Adapt the application:– Use an appropriate (joint source-channel) video codec
• very low bit rate with high average PSNR and acceptableframe rate (subband based, MPEG-4, H.263L etc.)
– Use a CBR voice connection (e.g GSM speech CODEC)
Adapt the communication layer:– Carry out appropriate resource management (bandwidth
reservation, allocation and utilization)– Implement a “smart” bandwidth partitioning strategy
– Implement a complimentary MAC protocol
August 8, 1997 Victor Bahl
%DQGZLGWK�5HVHUYDWLRQ��8WLOL]DWLRQ��DQG�3DUWLWLRQLQJ
QuestionCan performance guarantees be provided to VBR video withoutsignificantly under-utilizing the bandwidth?
AnswerYes! Use Intra-frame statistical multiplexing
August 8, 1997 Victor Bahl
6WDWLVWLFDO�0XOWLSOH[LQJ��&RQQHFWLRQ�/HYHO�DQG�6\VWHP�/HYHO
Text
VideoSource
Region 1 /Subband 1
Region 2 /Subband 2
Region n /Subband n
MUX
MUX
Audio Source
Data Source
Laptop
Connection-level Multiplexing
System-level multiplexing
Laptop
LaptopLaptop
August 8, 1997 Victor Bahl
(IIHFW����,PSURYHG�%DQGZLGWK�8WLOL]DWLRQ
Primary Region Usage
0
5
10
15
20
25
30
0 125 250 375 500 625 750 875
Time (Seconds)
Tot
al B
its G
ener
ated
(M
bits
)
Statistical Multiplexing within a Frame
0
0.2
0.4
0.6
0.8
1
0 3 6 9 12 15 18 21 24 27
Time (Seconds)T
otal
Bits
(M
bits
)
Cumulative Arrivals Reserved Bandwidth
Predicted Demand
Actual Demand
Bandwidth usage with and without intra-frame statistical multiplexing
August 8, 1997 Victor Bahl
%DQGZLGWK�3DUWLWLRQLQJ���6\VWHPV�3HUVSHFWLYH�
Possible Strategies:+ Complete Sharing
+ Complete Partitioning (or Mutually Restricted Access)+ Partitioning with Priority Borrowing (or Partial sharing)
Packet Voice
DataPacket Video
Packet Voice
DataPacket Video
Packet Voice
DataPacket Video
Strategy 1: Complete Sharing (CS) Strategy 2: Complete Partitioning (CP) Strategy 3: Priority Sharing (PS)
August 8, 1997 Victor Bahl
6KDULQJ�6WUDWHJLHV����3URV�DQG�&RQV
CP CS PS• Blocking level for
each class easy toadjust
• Blocking level amongclasses notadjustable
• May be tunable tovariable blockingrequirements
• Complete Protectionfrom overload ofother classes
• No protection fromoverload of otherclasses
• May offer protectionagainst overloadfrom other classes
• Requires too muchbandwidth
• Better Bandwidthusage than CP
• More efficientbandwidth usagethan CP and CS
August 8, 1997 Victor Bahl
3URSRVDO����3ULRULW\�6KDULQJ�ZLWK�5HVWULFWLRQV��365�
Packet VoiceData
Packet Video
Contend Reserved
Moving boundary Moving boundary
Fixed boundary
Data Voice Video-Dynamic
Video-Static
Data - BP BP BPVoice B - B BP
Video-Dynamic B BP - BP
Video-Static X X X -
B --- Borrow BP --- Borrow, but Preemption possible
RequesterReserved
August 8, 1997 Victor Bahl
&RPSDULVRQ�RI�%DQGZLGWK�3DUWLWLRQLQJ�6FKHPHV
Voice + Data Traffic
Voice Traffic Load = 30 ErlangsData Traffic Load = 10 Erlangs
Video Traffic
Compression = Region-based H.263Avg. Frame Rate = 7.5 HzAvg. Video Bit rate = 24 kbpsImage Dimensions = QCIF (176 x 144)# of Regions/Image = 5Peak for primary Region = 20 kbpsAvg. PSNR = 30.4 dB
Bandwidth Partitioning and VBR Video
0
5
10
15
20
25
30
35
5 25 45 65 85 105
125
Offered Video Traffic Load (Erlangs)
Ave
rage
PS
NR
PB
PSR
CS
August 8, 1997 Victor Bahl
&RPSDULVRQ�RI�%DQGZLGWK�3DUWLWLRQLQJ�6FKHPHV
voice = 60 erlangs
data = 10 erlangs
Bandwidth Partitioning and VBR Video
0
5
10
15
20
25
30
35
5 25 45 65 85 105 125Offered Video Traffic Load (Erlangs)
Ave
rage
PS
NR
PB
PSR
CS
August 8, 1997 Victor Bahl
2SWLPDO�%DQGZLGWK�3DUWLWLRQLQJ
ProblemCan performance guarantees be provided to VBR video withoutsignificantly under-utilizing the bandwidth and can this be done inconjunction with minimizing the blocking probability for voice and datatraffic?
AlternativelyHow do we decide, how much of the available bandwidth should beallocated to each traffic class?
Solution Allocate bandwidth so the maximum call blocking probability is minimized
August 8, 1997 Victor Bahl
0LQLPL]H�WKH�0D[LPXP�%ORFNLQJ�3UREDELOLW\
Problem Statement
Given the aggregate load Di for each traffic type Ti and the total systembandwidth B, determine the allocated transmission capacity for each traffictype such that
Θ = ≥max ( ( ) )i
i iP d t B is minimized
B Bii
N
=∑ =
1subject to where d ti ( ) is the instantaneous demand
August 8, 1997 Victor Bahl
Theorem
Let X1 , …, XN be independent random variables taking theirvalues from the interval [0, 1]. Their probability distributions areotherwise arbitrary and not necessarily identical. Setand then for any the following estimation holds:
Furthermore, this estimation is best possible in the following sense:For any fixed and for any fixed D and C with there existinfinitely many counter examples for which the reverse holds.
X Xii=∑
D E X=
C D≥
( )P X CD
Ce
CC D≥ ≤
−
C D≥ε > 0
0LQLPL]HΘ = ≥max ( ( ) )
iP di t Bi
August 8, 1997 Victor Bahl
0LQLPL]H
New Problem Statement
Given the aggregate load Di for each traffic type Ti and the totalsystem transmission bandwidth B, determine the allocatedtransmission capacity for each traffic type, such that
is minimizedΘ =
−max
i
i
i
B
B DD
Be
i
i i
subject to B Bii
N
=∑ =
1
Θ = ≥max ( ( ) )i
P di t Bi
August 8, 1997 Victor Bahl
%DVLV�IRU�'HWHUPLQLQJ�%L�V
Can prove: Allocation of Bi is asymptotically optimal if and only if
Thus the problem is solved by letting,
and iteratively searching for a value for whichholds within a given error bound
D
Be
D
Be
B
B D N
N
B
B DN
N N1
1
1
1 1
= =
− −K holds !
D
Bei
i
B
B Di
i i
=− σ
0 1< ≤σ B Bii
N
=∑ =
1ε > 0
August 8, 1997 Victor Bahl
%DVLV�IRU�'HWHUPLQLQJ�'L
Method
If Xi is the size of ith frame (or region or subband) in a video sequencewhich has n frames, and if we define:
then by letting and deriving
we can calculate the mean and determine Di
where M is the # of video connections to be supported
( )Y X X Xn n= max , ,1 2 K
( )f yYnY Y
nn=
→ ∞lim
D Mi y= × η
ηy
August 8, 1997 Victor Bahl
6WDWLVWLFDO�&KDUDFWHUL]DWLRQ��9%5�9LGHR�
-0.010
0.010.020.030.040.050.060.070.080.09
1 11 21 31 41 51Size (Kbits)
Pro
bab
ility
Den
sity
Observed Normal
Lognormal Gamma
Weibull
Bi-Normal distribution“Amadeus”
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
Size (Kbits)
Pro
bab
ility
Den
sity
Inter-Frame Coding“Headline News”
Gamma
Lognormal
Observed
Avg. PSNR = 30.44 dB
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
1 16 31 46 61 76 91
Size (Kbits)
Pro
bab
ility
Den
sity
Intra-Frame Coding“Headline News”
Avg. PSNR = 31.52 dB
QQ plot for “Headline News" (JPEG)Using Weibull Distribution
0
10
20
30
40
50
60
0 10 20 30 40 50 60
Normal Quantile
Ob
serv
ed Q
uan
tile
August 8, 1997 Victor Bahl
6WDWLVWLFDO�&KDUDFWHUL]DWLRQ��9%5�9LGHR�Mean Square Error (inter-frame compressed, CNN sequence)
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
0 8 16 24 32 40 48
Time (Seconds)
Mo
del
ing
Err
or
(Mb
its)
Captured & Compressed @ 15 fps
-0.05
0
0.05
0.1
0.15
0.2
0.25
5 10 15 20 25 30 35 40 45 50
Image Frame Size (Kbits)
Pro
bab
ility
Den
sity
Captured & Compressed @ 5 fps
-0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
0 5 10 15 20 25 30 35 40 45 50
Image Frame Size (Kbits)
Pro
bab
ility
Den
sity
Avg. Mean Square Error
Normal = 0.042793Gamma = 0.041567Lognormal = 0.037322Weibul = 0.070562
mean = 14.62par = 5.83cov = 0.57avg. dev.= 6.69
mean = 5.17par = 9.89cov = 1.17avg. dev.= 4.12
August 8, 1997 Victor Bahl
'HWHUPLQLQJ���������������DQG( )f yYn ( )Di yη
Coding Request Video Conference Video Technique
High Capture Low Capture High Capture Low Capture Rate Rate Rate Rate
High Action Low Action High Action Low Action Low Action Low Action
Intra-Frame Normal Normal/Gamma Normal/Weibull Weibull Bi-Normal Bi-Normal
Inter-Frame Bi-Normal Gamma/Lognormal Normal Gamma Gamma-Pareto Lognormal
Results from distribution-based modeling of single VBR video sources
F y e yYy u( ) exp[ ],( )= − − ∞ < < ∞− −α
( ) ( )F x exg x= − −1 ( )g x is a increasing function of X
Notice, most video frame distributions can be described as:
where
then we can prove,
August 8, 1997 Victor Bahl
'HWHUPLQLQJ���������������DQG
l Thus knowing the mean can be computed as:
where and are the location and scale parameters
l For Voice connections:
where C is CBR (e.g. GSM - 13 KHz, DECT - 32 KHz, …) and M isthe estimated number of voice connections
Note: Can use Di s for Network Capacity Planning !
( )F yY
ηαy u= +
0 577.
u ( )α > 0
D Ma = × C
( )f yYn ( )Di yη
August 8, 1997 Victor Bahl
7KH�6PDUW�$OORFDWH�$OJRULWKP
Begin
End
l u= =0 1,
σ =+l u
2
( )~B Bii
N= =∑ σ1
~B B< − ε
~B B>
u = σ
( ) ( )B B BN
BN1 1
= =σ σ, ,K
l = σ
Midpoint of [ ]l u,
Initialize
Yes
No
No
Compute TotalCapacity
Check if Capacityis overused
Check if Capacityis underused
Compute ( )Bi
σ
Stricterbound
Ease upon bound
August 8, 1997 Victor Bahl
�LOOXVWUDWLYH�([DPSOH
Assume TDM-based system with 99 BBU
Let the aggregate average demands beD1 (voice) = 20 Erlang and D2 (video) = 40 Erlangs
Then the Load Proportional approach givesB1 = 33, B2 = 66, with Erlang’s blocking probability = 1 %
Smart Allocate algorithm givesB1 = 37 and B2 = 62, blocking probability = 0.57%
Improvement of 43% !
August 8, 1997 Victor Bahl
&RQFOXVLRQ��$GYDQWDJHV�RI�365���6PDUW�$OORFDWH�$OJRULWKP
l Accommodates all traffic classes without shutting out any single one
l Provides QoS guarantees for on-going VBR real-time videocommunications. This guarantee does not come at the expense ofbandwidth and other traffic classes.
l Robust and insensitive to statistical assumptions. Does not requiredetailed knowledge of traffic, only aggregate average values (detailedstatistical information is typically unavailable)
l Allocation based on minimizing a bound on the blocking probabilitiesthat is proven to be asymptotically optimal - significant as it signifies thatfor large number of systems it is sufficient to know aggregate flow rates.
August 8, 1997 Victor Bahl
&XUUHQW�3URMHFWV�DW�0653HULSDWHWLF�&RPSXWLQJ�IRU�WKH�1H[W�0LOOHQLXP
l Ad-Hoc multi-hop home networking– Protocols, Routing algorithms, architecture, roaming, locating, performance etc
l Hardware for hand-held communicators– RF issues, form factor, capacity, capability, user requirements, user interface etc.
l QoS in mobile multimedia– Content-Sensitive Video Coding (beyond MPEG-4), Resource assignment and
management, etc.
l Operating system support for mobile users– caching, hoarding, prediction, data management, disconnected operations etc