investigating wireless systems two case studies
DESCRIPTION
Investigating Wireless Systems Two Case Studies. Roch Guerin University of Pennsylvania. Starting Point. Wireless means no wire… Flexibility in using and allocating spectrum resources No wire means lots of unpredictable interactions - PowerPoint PPT PresentationTRANSCRIPT
Investigating Wireless SystemsTwo Case Studies
Roch Guerin
University of Pennsylvania
2
Starting Point
• Wireless means no wire…– Flexibility in using and allocating spectrum resources
• No wire means lots of unpredictable interactions– User interferences, fluctuations in channel quality, e.g.,
fading, etc.
• What is the trade-off between flexibility and predictability?– A common theme at the “physical layer” (MIMO,
OFDM, etc.)– But how does it play out at the link/network layer?
3
Broad Problem Setting
• Multiple distinct channels, e.g., different frequencies, spreading codes, etc.
• One or more user seeks to transmit data• How should data be transmitted to maximize “performance?”
– Which user transmits when on what channel– What definition for performance (system centric versus user centric)
• Performance factors– Channel quality– Channel access schemes– Channel allocation and packet transmission strategies– User traffic patterns– User level coding for resiliency
4
Not an Easy Problem in General…
• M users and C (slotted) channels– An idle user becomes active with probability p and
remains active for k1 slots– Users can direct any transmission to any channel (no
channel switching cost)– Basic slotted Aloha MAC and error-free channels
• What’s the best assignment of users to channels– If CM, then obviously one user per channel is best– If CM, the answer depends on p and k
– Analysis is possible based on standard Markov chain formulation
• Nothing sophisticated, but complex state enumeration and bookkeeping
5Load
Thr
ough
put
Impact of Channel Allocation (k=1)(M,C)
(2,1)
(4,2)
(6,3)
(8,4)
(16,8)
(32,16)
Sharing pays only at high loads
M M =2
M =32
6
Our Focus• Two case studies within the broad framework we just
outlined• Case study 1: Controlled diversity
– Channels: Variable quality with known open-loop “statistics”– Channel access: No interferences/collisions between users– Performance goals: Coding and transmission policies to improve
individual user throughput• Case study 2: Distributed diversity
– Channels: Variable quality with closed control loop to equalize channels across users
– Channel access: A CDMA setting with “controlled” interferences between users
– Performance goals: Transmission policies that optimize trade-off between system throughput and transmission flexibility of individual users
7
Generic Outline for Both Case Studies
• Problem justification– Why someone else should care about it…
• Problem formulation– What parameters
• Prior work– What others have done
• Problem solution– What analytical tools and techniques
• Solution evaluation– Methodology and scope
8
Our Focus• Two case studies within the broad framework we just
outlined• Case study 1: Controlled diversity
– Channels: Variable quality with known open-loop “statistics”– Channel access: No interferences/collisions between users– Performance goals: Coding and transmission policies to improve
individual user throughput• Case study 2: Distributed diversity
– Channels: Variable quality with closed control loop to equalize channels across users
– Channel access: A CDMA setting with “controlled” interferences between users
– Performance goals: Transmission policies that optimize trade-off between system throughput and transmission flexibility of individual users
9
Case Study 1
• Should I do this? • Or that?
• And if that, how and when?
10
Why the Question?
• Mobile devices are increasingly powerful– They run pretty much the same range of applications as
wired devices– Many of these applications are performance sensitive
• Wireless resources are subjected to random fluctuations in quality that are outside the control of users and network providers alike
• So it’s worth exploring what we can do to make performance more predictable to users in spite of our limited control on the resources they access
11
Why the Approach?
• Spreading transmissions across multiple channels allows us to– Avoid being stuck with a really bad channel– Decrease the effective length of error bursts, which can
facilitate recovery (fewer consecutive lost packets)
• Potential for improvements, therefore, arises from– A higher probability of successful message transmission– The ability to lower coding overhead
• The questions are then how to realize the best possible improvement, and how big it is
12
System Overview• One user, C channels (remember perfect channel access…)
– User wants to transmit messages (data blocks) of size k packets
• Channel model is “known” – e.g., Gilbert-Elliot model– Independent channels with known statistics
• Reliable transmissions through packet-level code– Add redundant packets to achieve target probability of successful
message delivery Pmin
• Policy distributes packet transmissions across channels – Deterministic and probabilistic policies– No channel switching overhead
• Performance measure: Effective Rate (ER)– Number of messages successfully delivered per unit of time (unit
of time is packet transmission time)
13
Channel Model• Two-state Markov chain with Good and Bad states
– Packets are lost when channel is in bad state
– Long-term error rate
– Expected burst length
• More complex channel models can be constructed using higher order Markov chains– Increased computational complexity (of transmission policies)
G BPe
1-Pb
1-PePb
be
e
PP
PLTER
1
bPEBL
1
1
14
Transmission Model
• Fixed size message consisting of k packets
• Messages sent using (N,k) diversity code– Corrects all patterns of i N-k erasures
(erroneous or lost packets)
– N ≥ k is chosen to realize Pmin
• Policy A selects channel for each packet transmission
15
Transmission Policies
• Probabilistic policies– Before each packet transmission select channel
i, 1i C, with probability pi
– Policy specified by p = [p1 p2 … pC]
• Deterministic policies– For N-packet messages, pre-determine the
channel ci that packet i, 1i N, is to be sent on
– Schedule S=[c1, c2,…,cN] specifies policy
16
Performance Metric
• Identify policy A and code length N that maximize
where is the probability of successful message transmission given N and A, and N is the smallest value that satisfies Pmin
• For two policies A and B, the relative gain in ER of using B over A is given by
N
kNPkkNER
Asucc
A
),(),(
),(
),(),(),(
AAA
AAABBBER kNER
kNERkNERBAG
),( kNP Asucc
17
Related Works1. L. Golubchik, J. C. Lui, T. Tung, A.L. Chow, W.-J. Lee, G. Franceschinis,
and C.Anglano, “Multi-path continuous media streaming: What are the benefits?” Performance Evaluation, Vol. 39, Sept. 2002.
2. A. Tsirigos and Z. Haas, “Analysis of multipath routing-Part I: The effect on the packet delivery ratio.” IEEE Trans. Wireless Commun., Vol. 3, No. 1, Jan. 2004.
3. B. Abdouni, W. Cheng, A. L. Chow, L. Golubchik, W.-J. Lee, and J. C. Lui, “Multi-path streaming: Optimization and evaluation.” Proc. MMCN'05, San Jose, CA, Jan. 2005.
4. E. Vergetis, R.Guerin, and S. Sarkar, “Improving performance through channel diversity in the presence of bursty losses.” Proc. ITC 19, Beijing, China, Aug. 2005.
5. E. Vergetis, R. Guerin, and S. Sarkar, “Realizing the benefits of user-level channel diversity.” ACM Computer Communication Review, Vol. 35, No. 5, Oct. 2005.
6. E. Vergetis, E. Pierce, M. Blanco, and R. Guerin, “Packet-Level Diversity: From Theory to Practice. An 802.11-based Experimental Investigation.” Proc. ACM MOBICOM 2006, Los Angeles, CA, Sept. 2006.
18
Identifying Optimal PoliciesProbabilistic Policies• Calculate PA
succ(N,k) given the channel characteristics– Recursive solution– 4-state Markov Chain for
two independent GE channels
– For C independent channels, you end-up with a Markov chain with 2C states
• Search through all policies to find optimal selection
Deterministic Policies• Deterministic schedule allows
each channel to be viewed independently– Compute statistics of the
associated embedded Markov chains (one for each channel)
• Total number of errors is sum of independent random variables (number of errors when using each channel)– Use convolution to compute
overall probability of success
• Search through all policies to find optimal selection
19
Probabilistic PoliciesTwo Independent GE Channels
• Two GE channels give rise to a 4-state Markov Chain– Denote the stationary
probability of state i as i
• Let be the conditional probability of m errors in n transmissions under policy A, given that the initial channel state was i and the ending state was j
• We then have
),( nmP Aij
kN
m i j
Aiji
Asucc NmPkNP
0
4
1
4
1
),(,
20
Recursive Computation Procedure
• For all n = 0,1,… and m = 0,1,…,n and for all i,j{1,2,3,4}, we have
where Pkj is the transition probability from state k to state j
• Initial conditions are defined as
4
1
4
1
}{)1,1(
}{)1,(),(
k
Akj
Aik
k
Akj
Aik
Aij
jPPnmP
jPPnmPnmP
, is stateerror
is stateerror no
. and all for
otherwise; 0 and if
00),(
,1)0,0(
mnmnmP
jiPA
ij
Aij
21
Deterministic Policies• For each of the N packets, specify the channel used
– There are CN such policies…
• Focus on round-robin policy– Maximizes return time to channel, i.e., every C slots.
Ceb
eb
bstepCb
Ceb
eb
estepCe
PPPP
PP
PPPP
PP
)(11
11
)(11
• Let vi be the number of errors when using channel i– Distribution of vi is easy to calculate via a recursion
• Use convolution to calculate the distribution of
V= v1 + v2 + …+ vC and the performance of any (N,k) code.
22
Computational Challenges
• We have computational procedures to explore the space of possible policies, but while identifying optimal policies is feasible, it can be complex
• No computationally tractable “closed-form” solutions– Caused in part by the discrete nature of the problem– No consistent behavior of optimal policy
• Identical channels need not be used equally• Bernoulli channels not always preferred over burstier channels• More channels does not always improve performance
• What do we do next?– Look for “heuristics” to identify what policies are good, when
23
Methodology and Results Summary
• Step 1: Initial information gathering– Evaluate a broad range of channel combinations
• When does using multiple channels help?• Common characteristics of the optimum policy?
• Step 2: Data analysis and further evaluation– Most scenarios that yield “meaningful” (≥ 30%)
improvements use all channels roughly equally– What channel combinations give rise to such behavior?
• Concept of equivalent channels
• Step 3: Distilling a simple heuristic to quickly– Identify when and how using multiple channels helps
24
Step 1• All possible 1540
channel pairs between 55 different channels– LTER [1%, 9%]– EBL [1.01, 20] pkts
• Different combinations of values for k and Pmin
• Initial findings– Max. benefits when
channels are used roughly equally
– Necessary but not sufficient
25
The Cost of Using Channels Equally
• Define the loss in performance gain as
L = Gopt – Gequal
Average 1.28%
Std. Deviation 2.71%
Minimum 0.00%
Maximum 15.72%
Median 0.03%
L 5% 81.62% of cases
L 10% 98.77% of cases
L 15% 99.04% of cases
L 20% 100% of cases
26
A Different Look at the Data
27
Equal Channel Use Focus on Deterministic Policies
Easier and better
28
Step 2
• Understanding when channels are used equally– Holds for identical channels in most settings– Any other scenarios?
• Classifying channel combinations– Channels are used equally under the optimal policy– Channels have identical individual performance (ER)– Channels yield the highest gain when used together
• Simple test to identify “good” combinations – All three above perspectives give similar answers Look for combinations of channels with similar
performance
29
“Equivalent” Channels – (1)
30
“Equivalent” Channels – (2)
• Optimal policy remains close to 0.5 for “equivalent” channels
31
Step 3 - Simple HeuristicGiven C channels• Identify subsets of ~ equivalent channels
– Compute ERi, for each channel i, 1 i C– Group channels into |E| “equivalence classes”
• For each equivalence class eE with ne channels– Compute achieved by cycling through all channels
– If , use all channels
– Else use channel
– Set
• Pick equivalence class
)(max
)(max
1
1
)(equal
ini
ini
e
ER
ERER
e
e
)(equal
eER
)(argmax1
)(i
ni
e ERe
),max( )()(
equal)(
eERERER ee
)(argmaxˆ )(e
eERe
32
Exploring Further When Using Multiple Channels Helps
• Three parameters of interest:
1. Channel characteristics, i.e., EBL and LTER
2. Performance target Pmin
3. Number of channels available
• Focus on the case of two identical channels
33
Impact of Channel• The gain is biggest for bad channels
34
Impact of Pmin
• The more stringent the performance, the greater the gains
• Relative burstiness of channel plays a major role
No gain
Rapid transition to 50/50 policy
Large gains until relative burst size decreases N
35
Impact of Number of Channels
Biggest bang for the buck with just a few channels
36
Shifting Focus• Better performance is great, but
– How robust are the improvements?• Sensitivity to measurement inaccuracies, non-stationary
channels, etc.– Are we optimizing ourselves into a corner?
• Explore sensitivity to errors in channel estimates– Different channel statistics
• EBL and LTER• Distribution of duration of error bursts (not a GE channel!)
• Can we trade-off optimality for robustness of solution
– Already do this to some extend with round-robin policy
37
Sensitivity to Channel Quality
• Three users, three channels– Two scenarios: (1) each user is assigned one channel; (2) all three users
share the three channels– Both EBL and LTER are progressively made worse
• First on only one channel (left), then on all three channels (right)
• Better performance also comes with mostly greater robustness!– Exception in the single bad channel case, when both EBL and LTER are
over 40% worse
38
Performance vs. Robustness (1)
• Explore trade-off by varying the code length N – Bigger N greater robustness, but lower performance gain– Initial focus on basic channel statistics (LTER and EBL)
System Performance gain over the one channel system
Increase in both LTER and EBL before target Pmin is violated
One channel (N=19) 0% 2%
Three channels (N=15) 27.6% 16%
Three channels (N=16) 20.7% 37%
Three channels (N=17) 14.2% 63%
Three channels (N=18) 8.2% 92%
Three channels (N=19) 2.7% > 100%
39
Performance vs. Robustness (2)
• This time we vary the channel model (target Pmin=0.97)– Ever burstier channel (variance of error burst )
Variance Multiplier
One channel Three channels
N = 19 N = 15 N = 16 N = 19
ER Psucc ER Psucc ER Psucc ER Psucc
Original 1.534 0.971 1.956 0.978 1.850 0.987 1.574 0.997
x 0.25 1.555 0.985 1.947 0.973 1.840 0.982 1.574 0.997
x 0.5 1.547 0.980 1.942 0.971 1.837 0.980 1.568 0.993
x 1 1.538 0.974 0 0.968 1.83 0.976 1.562 0.989
x 2 0 0.963 0 0.962 0 0.968 1.552 0.986
x 4 0 0.961 0 0.949 0 0.957 1.538 0.974
x 8 0 0.953 0 0.941 0 0.949 0 0.966
40
Resting on our Laurels – Or Not?
• Theory tells us that we can have our cake and eat it!– Both better and more robust performance
• But the theory is rife with holes and assumptions– Independent, stationary channels, with known statistics– No impact of user transmissions on channel statistics– No channel switching overhead
• Putting it all to the harsh test of reality– 802.11b environment– Standard end-systems (PCs) without precise control of
transmission timings
– If it survives “that” then maybe there is some hope…
41
Surviving 802.11
• Performance is all over the place…
• There is no “average” 802.11 channel– Stationary GE model not particularly accurate– Significant time-of-day and location dependent variations
• Wild fluctuation across 10 minute intervals– LTER can range from 0.01% to 70%– EBL varies between 1 and 40 packets
• Actual error bursts between 1 and several hundred packets
• Similar observations made by others
• Does not bode too well for the “survival” of our theory– Nevertheless
42
Experimental Setup• Two 802.11b Access Points (APs)
– Intel StarEast board, with one miniPCI NIC each– External omni-directional antennas– Assigned “non-overlapping” frequency bands– Located ~1m from each other– Transparent logging of all incoming packets on both systems– Within reach of other APs interfering in all 11 frequency bands
• Sender– Standard laptop with two NICs
• One external PCMCIA NIC, and one internal miniPCI NIC• Linux operating system• Transmission speed set at 2Mbps
– Located between 2m and 10m away from the two APs• Maintains association with both APs• Line-of-sight (LoS) as well as non-LoS (indoor wall) transmissions
43
Some Other Implementation Issues
• Impact of 802.11 operation– RTS/CTS handshake before transmissions
[Disabled - Large RTSThreshold value]– “Feedback” mechanism: ACK packets
[Disabled - Broadcast packets ]– Channel access control (untouched)
• Sensing and exponential backoff• Inter-frame spaces (SIFS, DIFS, etc.)
• Processor and OS overhead vs. transmission speed of the NICs– Where is the bottleneck and how does it affect
transmission timings?
44
Transmission Timing Scenarios
• Single channel
• Interleaving on one channel
• Perfect timing on two channels
• Bandwidth limited on two channels
• Processor limited on two channels
• Interleaving on two bandwidth limited channels
NIC1
NIC1
NIC1
1 2 3 4 5 6 7 8
1 2 3 41’ 2’ 3’ 4’
NIC2
1
2
3
4
5
6
7
8
NIC1
NIC2
1
2
3
4
5
6
7
8
NIC1
NIC2
1
2
3
4
5
6
NIC1
NIC2
1
2
3
4
5
6
7
81’ 5’
2’
3’
4’ 6’
7’
8’
(S0)
(S1)
(S2)
(S3)
(S4)
(S5)
45
Experimental Evaluation
• Generate extensive sets of traces– “Continuous” transmissions on both NICs
• Traces of received packets (1,000 bytes) recorded at each AP
– Different configurations• Sender location, time-of-day, selection of frequency bands
– With and without interferer in “intermediate” band• Test for channel correlation
• Post-process traces to emulate different settings– Different combinations of system parameters: N, k, Pmin – Various packet inter-leaving strategies, transmission
timing policies, channel switching overhead, etc.
46
The Net of It(see MOBICOM Paper for Details)
• Generic findings– Channel correlation was not found to be a major issue, at least
when using non-overlapping channels– Precise timing of transmissions does not appear to have a major
impact on performance– With current technology, dynamic channel switching mandates the
use of multiple radio cards• Cannot compensate for it through smart scheduling
• Performance/robustness findings – Some benefits remain in spite of 802.11 channel “characteristics”
IF channel characteristics are known• Unlikely to be of much use in practice given the erratic nature of the
802.11channel– For unknown channels, meaningful benefits remain ONLY when
some of the channels are really bad• Benefits are more as a performance stabilizer than an enhancer
47
Stabilizing Performance
• Two channels:– LTER1 ~ 11.4%
– EBL1 ~ 10 pkts
– LTER2 ~ 29.2%
– EBL2 ~ 11 pkts
• Measure ER over a 200 messages sliding window– Mean value improves
by 6%/30%
– Variance decreases by 60%/90%
Channels 1+2
48
Wrapping Up Case Study 1• When and how can using multiple wireless channels be of
benefit?– Assumed no user interference (centralized control) and no channel
feedback (open-loop)– Relied on packet-level diversity coding– Ignored overhead in changing channel
• Analytical tools– Basic Markov chain analysis (stationary probabilities, recursions)– Elementary probability (convolution, etc.)
• Main results– Cycling in a round-robin manner across channels is close to being
optimal in most cases– Better performance is reasonably robust to errors in assumptions– Even when most of the above assumptions don’t hold, e.g.,
802.11, there are advantages to spreading transmissions across channels
49
Our Focus• Two case studies within the broad framework we just
outlined• Case study 1: Controlled diversity
– Channels: Variable quality with known open-loop “statistics”– Channel access: No interferences/collisions between users– Performance goals: Coding and transmission policies to improve
individual user throughput• Case study 2: Distributed diversity
– Channels: Variable quality with closed control loop to equalize channels across users
– Channel access: A CDMA setting with “controlled” interferences between users
– Performance goals: Transmission policies that optimize trade-off between system throughput and transmission flexibility of individual users
50
Case Study 2• What is the cost of
going from this?• To that?
• And what’s the
best trade-off?
51
Why the Question?
• Mobile devices are increasingly powerful– They run pretty much the same range of applications as wired
devices– These applications exhibit a broad range of resources and
performance requirements
• Tight centralized control can optimize overall system performance, but may be a very poor match to individual user needs
• So it’s worth exploring the trade-off associated with giving user more independence in their transmission decisions and the impact this has on overall system throughput
52
System Overview
Internet
TelephoneNetwork
Base StationController
MobileSwitching
Center
InternetGateway
Base Station
53
Our Focus
54
Overview of CDMA Uplink
• CDMA uplink is interference limited– Each user has a spreading “orthogonal” code
• Allows simultaneous transmissions
• However, users interfere due to multi-path effects
• Users can select among multiple (discrete) transmission rates– Control loop based on pilot signal equalizes channel
among users
– Transmitted power is proportional to pilot strength AND selected rate
55
Uplink Operation
• Pilot Pi transmitted by device i=1,...,n+1– Pilot signals are power controlled by BS to all be received
with the same target SINR 1/Ф
• Giloss : Path loss; θPilot: Orthogonality factor; σ2 : Noise
• User i transmit power = Pi · TxT2P[R]
– R : Target data rate from discrete set – TxT2P[R] : Proportionality factor relative to Pilot
• User spends TxT2P[R] power tokens to transmit at rate R
1,,1,1 2
2
nin
PGPG
PG
Piloti
iloss
ijj
jlossPilot
iiloss
56
Sample TxT2P[R] Values
Target Data Rate TxT2P[R] dB
0 -∞
9.6 kbps 4.5
19.2 kbps 6.75
38.4 kbps 9.75
76.8 kbps 13.25
153.6 kbps 18.5
57
CDMA Uplink Interference Model
Pilotij
j
iiii
iiii
Di
i
ijj
jD
jloss
ii
Dilossi
ii
nRPTxT
RPTxTRGRSINR
RPTxTPRPR
WRG
RPG
RPGRGRSINR
2
2
2
,][2
][2)()(
][2)( andGain Processing :)(
factority orthogonal Data :,)(
)()()(
No Channel Effects(Perfect Power Control)
• Interferences from other users– The higher the rate a user chooses, the more
interference it creates!
58
Our Problem
1,,1,][2
][2)()(
2
niRPTxT
RPTxTRGRSINR
ijj
iiii
• Users make independent transmission and rate selection decisions– Greedy behavior by individual users can affect overall
performance
• What guidelines to mitigate negative impact of independent decisions
59
Related Works7. K. Kumaran, L. Qian, “Uplink Scheduling in CDMA Packet-Data
Systems.” Proc. INFOCOM 2003, San Francisco, CA, April 2003.8. R. Cruz, A. Santhanam, “Optimal Routing, Link Scheduling and Power
Control in Multi-Hop Wireless Networks.” Proc. INFOCOM 2003 , San Francisco, CA, April 2003.
9. P. Venkitasubramaniam, S. Adireddy, and L. Tong, “Opportunistic ALOHA and cross-layer design in sensor networks.” Proc. IEEE MILCOM, Boston, MA, October 2003.
10. P. Venkitasubramaniam, Q. Zhao, and L. Tong, “Sensor networks with multiple mobile access points.” Proc. 38th Annual Conference on Information Systems and Sciences, Princeton, NJ, March 2004.
11. X. Qin and R. A. Berry, “Distributed approaches for exploiting multiuser diversity in wireless networks.” Trans. Infor. Theory, vol. 52, no. 2, pp. 392-413, February 2006.
12. A. Sridharan, R. Subbaraman, and R. Guerin, “Distributed Uplink Scheduling in CDMA Networks.” Proc. Networking'2007, Atlanta, GA, May 2007. (Extended version – Sprint Research Report).
60
Our Initial Model• Homogenous, unconstrained users
– All users (n+1 users in a sector) employ the same policy– Users always have data and are able to transmit whenever the
policy schedules a transmission• Probabilistic On-Off transmission policy
– Transmit at rate R in a slot with probability p• Transmit power is therefore 0 with probability 1-p and
~TxT2P[R] with probability p• Simple but useful model
– Similar to Aloha– But with a contention model based on soft interferences (CDMA)
rather than “collisions”• Questions
– At what rate R should a user transmit?– How often (what p value) should a user transmit?
61
Revisiting our CDMA Uplink Model
• We have
• Achieved rate
• On-Off policy with parameters R and p
where for simplicity we have assumed SINRi(R)S0
][2,)(
)(2 ii
iji
iiii RPTxTK
K
KRGRSINR
K
npp
j
n
jS
WppC Pilotinj
n
j
,)1(
1)(ˆ
00
][)(ˆ,)(
,min0
iiii
ii CEpCRS
RSINRRC
62
Rate-SINR Model
0S
iR
SINR
Ach
ieve
d R
ate Linear Model
Bounded Model
No matter how good the channel, you cannot get more bits out than you put in…
63
Main Results
• There exists an optimal p* (maximizes )– If 1 then p*=1– If < 1 then p* < 1– In both cases p* satisfies the following equality
– With few (many) users, and/or low (high) target rate R, users should transmit (in)frequently
• Higher target rates always achieve higher throughput, i.e.,– In the absence of other constraints
)(ˆ pC
1*)1(
1*)1(*
1
0 pnpp
j
n
jjnj
n
j
212*21
*1 ),,(ˆ),(ˆ RRRpCRpC if
K
n Pilot
64
Impact of
65
Hybrid Slotted/CDMA
66
Distributed Control
• Token bucket mechanism available in EV-DO Rev. A and HSUPA to “control” device transmissions– Token bucket depth and token fill rate are
controlled by Base Station– A device needs TxT2P[R] tokens to transmit at rate R– Aimed at limiting peak and average power to satisfy
fairness and QoS constraints• Question: How does the presence of a token
bucket affect the choice of “good” transmission decisions by devices?
67
Accounting for Token Buckets
• Given a token bucket configuration (,)– What are the optimal p* and K values?
• Two-step formulation1. Account for impact of token bucket on transmission
decisions• Transmissions conditioned on having at least K tokens
2. Explore possible combinations of p and K values– Note that optimality of higher rates need not hold any more
because of token constraints (token efficiency)
68
Token Efficiency
• With 24 users transmission at 153.6kbps yields a higher throughput but a lower token efficiency than transmission at 76.8kbps
69
Impact of Token Bucket
Conditional Transmission Probability
Token Bucket parameters:
σ = 21.5dB; = 7dB
More frequent transmissions at 76.8kbps yield a better throughput because of higher token efficiency
70
Token Bucket Formulation
• Let p and ptok denote conditional and unconditional transmission probabilities– Token bucket evolution governed by simple Markov chain
– So that
• Optimal (K,p*) value satisfies non-linear program N1
0 K+1K1 σ-K σ 1-p
1-p1-p1-p1-p
p p p
11
10,11
1
tok
K
iitok ppp
jntok
jtok
N
jtok
Kppp
j
n
jp
S
WKpCKpC
)1(1
),(ˆ where),,(ˆmax10
,
71
Solving Program N1
• Step 1: For each value of K, solve the unconstrained problem, i.e., look for actual transmission probability pu*(K) that maximizes throughput– Solved based on value of δ and fact that
• Step 2: For each value of K compute p*(K) that satisfies
• Step 3: Choose the pair (K,p*(K)) that yields the largest throughput
1*)1(
1*)1(*
1
0 pnpp
j
n
jjnj
n
j
1
1
* 1)()(minarg)(*K
iiu
pKpKpKp
72
Analysis vs. Reality
Token Bucket: σ = 21.5dB; = 7dB
Rate
(kbps)
Analysis Simulations
(bounded rate model)
p*A C*
A p*sim C*sim Csim(p*
A)
76.8 1.0 26.4 0.35 17.84 16.56
153.6 0.21 42.9 0.25 10.63 10.59
• Expected inaccuracies because of bounded rate
– But actual impact on throughput is small
73
Related Results and Extensions
• Recent results– Similar results also hold for the bounded rate model– Characterized optimum centralized schedule
• A benchmark against to compare distributed policies• A combinatorial problem because of discrete rate values
• Extensions– Investigating the impact/use of token bucket for its
“original” purpose, namely, service differentiation• Rate vs. delay performance targets
74
Summarizing Case Study 2
• What is the impact on the throughput of CDMA uplinks of uncoordinated user transmissions?– Used simple probabilistic policies to probe the effect of distributed
transmission decision– Extended the investigation to account for the effect for token
constraints imposed by the base station to “control” device transmissions
• Analytical tools– Algebraic manipulations and basic real analysis– Optimization techniques and Markov chains
• Main results– Identified optimal transmission strategy function of system load– Results also hold in more realistic setting of bounded rate model