Upload
dangkiet
View
222
Download
0
Embed Size (px)
Citation preview
Matrix Analytic Methods -Some Real Life Applications
David LucantoniDLT Consulting, L.L.C
www.DLTconsulting.com
Queueing Theory and Network Applications, Korea University, 2011
Objectives of this TalkReview Matrix Analytic Methods pioneered by Marcel Neuts
and others
Present real life examples where Matrix Analytic Methods have been used successfully in industryCongestion due to Internet Dial UpApplication to Common Channel Signaling (SS7)Modeling WiFi Hotspot TrafficMultiplexing bursty traffic in an ATM or IP networkCall Acceptance for a VoIP SystemTransient Results and applications
Queueing Theory and Network Applications, Korea University, 2011
Objectives of this Talk (Cont’d)No new mathematical analyses
A detailed description of any of these models would take the whole talk
Will only give general description of model parameters
Detailed definitions and analyses are widely available in the literature
Will give some matrix results along with the corresponding scalar (exponential or Poisson) results
Queueing Theory and Network Applications, Korea University, 2011
Analysis of Matrix Analytic ModelsModels are analyzed by purely probabilistic arguments (e.g.,
Markov Renewal Theory)
Explicit expressions lead to very stable numerical algorithms for computation of performance metrics
Matrix analytic methods avoid
the complex analysis used in previous approaches
the necessity of numerically computing the roots of an analytic functionNumerical instability when roots are close to each other or multiple
Queueing Theory and Network Applications, Korea University, 2011
Phase-Type (PH) DistributionsWill only consider continuous-time distributions here
Phase-type distributions contain many well-known and popular distributions
Exponential distribution, (m=1)
Erlang distribution
Hyper-exponential distribution
Queueing Theory and Network Applications, Korea University, 2011
Phase-Type Distributions (Cont’d)Defined as the time till absorption in an (m+1)-state Markov
process with 1 absorbing state
Let T be the m×m matrix of infinitesimal rates between transient states, α, the vector of probabilities of starting in a particular transient state and T0 be the rates into the absorbing state
Let X be a phase-type random variable and F(x)=P(X≤x)
Then
Queueing Theory and Network Applications, Korea University, 2011
Assuring Emergency Services Access (e.g., 911) in the presence of long holding times
due to Internet dial-up calls
The ProblemNo dial-tone condition - “Customers hear thin air.”VERY SERIOUS – e.g., 90/day life threatening emergency calls . Priorities non-implementablePreliminary suspicion on maintenance activities.Could it be congestion ? Red flag – internet usage !Model solved using a finite quasi-birth-and-death process (QBD)
[Ramaswami, et al., 2005]
Queueing Theory and Network Applications, Korea University, 2011
Cable Telephony
Network Interface
Unit
Cable Modem
HDT
CMTS
Backbone
Backbone Transport OC-48+
CLASS -5
SWITCH
SS7
PSTNMaster Headend or
Primary Hub
(100K - 200K HHP) Each
Zone 1 Zone 2 Zone 3 Zone 5
Central Office
DWDM
DW
DM
Fiber Node 16 Per
Secondary Hub
Zone 4
Transport
AD
M
AD
M
GR-303 GR-303
DA
CS
IP
Proprietary Protocol
TDM/T-1
DHCP/DNS TFTP
Servers
EMS SystemsDOCSIS
Protocol
Queueing Theory and Network Applications, Korea University, 2011
Effect of Long Holding Times (Cont’d)
Preliminary Analysisproblem in evening hours 7 PM – 11 PMno discernible spatial pattern
Push-backsEngineering followed “standard” proceduresOnly 5-8% of calls are internet dial-upsField measurements “do not support” congestion hypothesis
Queueing Theory and Network Applications, Korea University, 2011
Phase Type Model(using the EM-algorithm) ; fit uses ~4M data points
Phase Type Fit
Probability Observed Fitted0.1 5.4 5.80.2 12 12.70.3 21 21.20.4 32.4 32.20.5 48 47.70.6 72.6 72.20.7 120.6 120.20.8 232.8 235.90.9 601.8 597.6
0.95 1237.8 1268.00.99 3952.2 4035.2
0.995 6074.4 7168.4
Queueing Theory and Network Applications, Korea University, 2011
The Length Biasing EffectRemaining holding time distribution: [1-F(x)] / μ
RESIDUAL DISTRIBUTIONh(x)=[1-F(x)]/μ
Probability Observed Residual0.1 5.4 40.80.2 12 116.10.3 21 238.00.4 32.4 423.70.5 48 703.00.6 72.6 1124.60.7 120.6 1809.70.8 232.8 3285.30.9 601.8 7028.2
0.95 1237.8 11073.80.99 3952.2 20489.2
Median 48 703.0Mean 297 2367.0
Queueing Theory and Network Applications, Korea University, 2011
Empirical Validation Empirical data on residuals (~11 K observations) vs.
Model
Queueing Theory and Network Applications, Korea University, 2011
Impact of Long Holding TimesInsensitivity relates only to blocking over an infinite horizon.
Heavy tails PERSISTENCE of congestion. So, short term performance is quite different.“Standard” procedures do not work !
Bad periods compensated by long very good periods. Good situation for control !!! (See the paper and patents for possible controls.)
Queueing Theory and Network Applications, Korea University, 2011
Several Other Real Life Applications
Excessive Link Oscillations in the Common Channel Signaling (SS7) Network [Ramaswami and Wang, 1993]Solved using Phase-type distributions
Modeling and characterization of large-scale Wi-Fi traffic in public hot-spots [Ghosh, et al., 2011]Solved using an M/G/∞ where G is ln PH
distribution
Queueing Theory and Network Applications, Korea University, 2011
The Batch Markovian Arrival Process (BMAP)
Natural generalization of the Poisson process
Contains many well-known arrival processes (in all cases, correlated batch sizes are easily handled)Poisson process (m=1)PH renewal processMarkov modulated Poisson process (MMPP)Superposition of BMAPs is again a BMAP
Useful for modeling arrivals to an internal node in a network
Queueing Theory and Network Applications, Korea University, 2011
The Batch Markovian Arrival Process (Con’t) Consider an m-state Markov process and assume that at
each transition of this process, there is a probability of a batch arrival that depends both on the state before and after the transition
Let Dn, n≥1, denote the infinitesimal rate of a batch arrival of size n, keeping track of the underlying states before and after the transition
The generating function D(z) is given by
Queueing Theory and Network Applications, Korea University, 2011
BMAP/G/1 Queue The BMAP/G/1 queue is extremely useful for modeling the
performance of broadband packet networks Traffic is bursty Service times are NOT phase-type
Asynchronous Transfer Mode (ATM) Fixed length cells result in deterministic service times
IP networks Finite packet sizes finite mixture of deterministic service
times
BMAP models very bursty traffic and is closed under superposition, i.e., modeling the output of nodes entering another node
Queueing Theory and Network Applications, Korea University, 2011
Busy Period of the BMAP/G/1 QueueLet h(s) be the Laplace-Stieltjes transform of the
service time distributionLet G(z) be the probability generating function of
the number of customers served during a busy period (keeping track of the arrival phase). Then we have
Queueing Theory and Network Applications, Korea University, 2011
Busy Period of the BMAP/G/1 Queue (Cont’d)
Define G to be G(1) and g to be the stationary probability vector of G
It is easy to show that g is also the stationary probability vector of the phase of the arrival process during idle periods
For M/G/1, the probability that the system is empty is 1-ρ; for BMAP/G/1, the probability that the system is empty and that the arrival process is in phase j is the jth element of (1-ρ)g
Queueing Theory and Network Applications, Korea University, 2011
Virtual Waiting Time of the BMAP/G/1 Queue
Let W(s) be the Laplace-Stieltjes transform of the virtual waiting time distribution (keeping track of the arrival phase)
Let g be the stationary probability vector of the stochastic matrix G
Then we have
Queueing Theory and Network Applications, Korea University, 2011
Stationary Queue Length at DeparturesLet Xn be the stationary probability that there
are n customers in the queue at departuresLet , then
Queueing Theory and Network Applications, Korea University, 2011
Computing these DistributionsWe compute these distributions by numerically
inverting the transforms [Choudhury, Lucantoni and Whitt, 1994]
These algorithms are for inverting multi-dimensional transforms which we use later for computing transient distributions
Queueing Theory and Network Applications, Korea University, 2011
Known Results for the BMAP/G/1 Queue
Stationary distributions [Lucantoni, 1993] Queue length distribution at arrivals, departures, arbitrary
time Waiting time distributions
Transient distributions (given the appropriate initial conditions) [Lucantoni, Choudhury, and Whitt,1994], [Lucantoni, 1998] P(system empty at time t) Workload at time t Queue length at time t Delay of nth arrival Queue length at nth departure P(nth departure occurs less than or equal to time x)
Queueing Theory and Network Applications, Korea University, 2011
Modeling the Input to an Interior Network Node
Consider the output from several nodes which all go to the same node
Ideal candidate for the MAP/G/1 queue
Queueing Theory and Network Applications, Korea University, 2011
Multiplexing Bursty Traffic in an ATM Network
ATM requirement is that the probability of blocking must be less than 10-9
Tails of distributions are the most important
performance measure
Superposition of 64 on-off sources each modeled as a 2-state MMPP
Exact Analysis
Far from Poisson
Exponential tail does not become relevant until blocking is 10-40
Queueing Theory and Network Applications, Korea University, 2011
Advantage of MAP/G/1 Approach over Matrix-Geometric Approach When service
time has Many Phases
We computed the tail of the waiting time distribution in an MMPP64/E1024/1 queue by an by a distribution that is close to an E1024 and matches the first three moments of the deterministic distribution
The matrix-geometric approach would require using matrices of order 65,536
The MAP/G/1 approach looks at the process embedded at departures; since we don’t have to keep track of the phase of service, the matrices are only of order 64
Queueing Theory and Network Applications, Korea University, 2011
Advantage of MAP/G/1 (Cont’d)
G satisfies
When H is E1024, the integral on the right hand side is computed by inverting a 64×64 matrix 10 times
Much easier than computing the matrix exponential
Queueing Theory and Network Applications, Korea University, 2011
IP Data Traffic is Bursty
Sample Measurements from a previous assignment
Sampled every second
Highly burstyMean = 720 pkts/sec
Queueing Theory and Network Applications, Korea University, 2011
0 1 2 3 40
500
1000
1500
2000
2500
3000
3500
4000
4500
5000Packet Counts
Hours
Pac
kets
Simulated Poisson Traffic
Poisson traffic looks bursty
However, note the scale of the y-axis
Queueing Theory and Network Applications, Korea University, 2011
0 5 10 15640
660
680
700
720
740
760
780
800Packet Counts - Poisson
Minutes
Pac
kets
Poisson Process Compared to Data
Queueing Theory and Network Applications, Korea University, 2011
0 5 10 150
500
1000
1500
2000
2500
3000
3500Packet Counts - Poisson
Minutes
Pac
kets
0 5 10 150
500
1000
1500
2000
2500
3000
3500Packet Counts - Measurements
Minutes
Pac
kets
Fit an D-MMPP to Data[Heyman, Lucantoni,
2003]Choose λ1 so that the
peak is at the 95th percentile of a Poisson distribution, i.e.,
Then
Queueing Theory and Network Applications, Korea University, 2011
Fit an MMPP to Data (Cont’d)Lower bound of data covered by λ1 is
Let this be the upper bound covered by λ2, i.e.,
Then
Continue…
Queueing Theory and Network Applications, Korea University, 2011
Fit an MMPP to Data (Cont’d)Compute the transition probabilities by
For performance computations convert to a continuous time MMPP (see paper)
Queueing Theory and Network Applications, Korea University, 2011
Fit an MMPP to Data
Q-Q plot: plots quantiles
Should be linear if both come from same distribution
Very good fit
Queueing Theory and Network Applications, Korea University, 2011
Simulated MMPP Traffic
Queueing Theory and Network Applications, Korea University, 2011
0 5 10 150
500
1000
1500
2000
2500
3000
3500Packet Counts - MMPP
Minutes
Pac
kets
0 5 10 150
500
1000
1500
2000
2500
3000
3500Packet Counts - Measurements
Minutes
Pac
kets
Another Example
Queueing Theory and Network Applications, Korea University, 2011
0 5 10 150
100
200
300
400
500
600Simulated MMPP Packet Counts
Minutes
Pac
kets
0 5 10 150
100
200
300
400
500
600Measurement Packet Counts
Minutes
Pac
kets
Call Acceptance for a VoIP System A startup company was offering VoIP service to clients by buying a fixed
bandwidth virtual circuit (“fat pipe”) through the Internet
Wanted to know how big the fat pipe should be to handle a certain number of voice calls and still meet the Quality of Service (QoS) e.g., tail of the delay distribution
Measurements were taken for IP voice calls for different codecs, e.g., G.723, G.729, etc., and different measurement intervals
We developed a program where the user inputs the file containing the measurements, the measurement interval and the target delay
Queueing Theory and Network Applications, Korea University, 2011
Call Acceptance for a VoIP System (Cont’d)We fit an MMPP to the measurement data and by solving
the delay using an MMPP/D/1 queue
If the computed QoS is greater or less than the targeted QoS, we adjust the service rate and re-compute; continue adjusting the rate using a binary search until the computed QoS is within a given tolerance
The program then outputs the size of the required fat pipe and the required buffer size
Queueing Theory and Network Applications, Korea University, 2011
Notes on Transient ResultsGeneralized every result derived by Takács for
the transient M/G/1 queue [Takács, 1962]
All results are direct matrix analogues of the scalar results
Using only probabilistic arguments, resulted in expressions that can be computed numerically without computing the roots of an analytic equation
Queueing Theory and Network Applications, Korea University, 2011
Sample Transient Result (emptiness function)Consider a MAP/G/1 queue (similar results hold for the
BMAP/G/1 queue)
Let Px0(t) be an m×m matrix where the (i,j) entry is the probability that the system is empty at time t with the arrival process is in phase j given the at time 0 the work in the system is x and the arrival process is in phase i
Let px0(s) be the Laplace-Stieltjes transform of Px0(t)
Queueing Theory and Network Applications, Korea University, 2011
Emptiness Function (Cont’d)Then
Queueing Theory and Network Applications, Korea University, 2011
Transient Performance MeasuresWe compute the virtual waiting time (work in
system) at time t for a given queue length at time 0 by numerically inverting the 2-dimensional transform, W(s1,s2)
Queueing Theory and Network Applications, Korea University, 2011
Sample Transient Results -Virtual Delay Distribution
Superposition of four identical two-state MMPP’s
Service times are distributed as E16
Utilization = 0.7queue length =
0 at time 0queue length =
32 at time 0
Queueing Theory and Network Applications, Korea University, 2011
Sample Transient Results -Unstable System
Utilization (ρ) = 2.0
Queue length = 0 at time 0
Queueing Theory and Network Applications, Korea University, 2011
Potential Use of Transient ResultsCall acceptance algorithms
Current call acceptance algorithms are based on stationary probabilities
Given the number of calls in the system, the time to reach a stationary distribution can be MUCH longer than the inter-arrival and inter-departure times
Call acceptance tables could be computed a-priori based on decision time intervals, the current state of the system, the average and peak bandwidth and burstiness parameter of the arriving call
Queueing Theory and Network Applications, Korea University, 2011
References1. V. Ramaswami, D. Poole, S. Ahn, S. Byers, and A. Kaplan, "Assuring emergency
services access: .providing dial tone in the presence of long holding time internet dial-up calls," Interfaces, vol. 35, pp. 411-22, 2005
2. V. Ramaswami and J. L. Wang, "Analysis of the Link Error Monitoring Protocols in the Common Channel Signaling Network," IEEE/ACM Transactions on Networking, vol. 1, pp. 34-47, 1993
3. Ghosh, R. Jana, V. Ramaswami, J. Rowland, and N. K. Shankaranarayanan, "Modeling and characterization of large-scale Wi-Fi traffic in public hot-spots," in IEEE INFOCOM, Shanghai, China, 2011
4. D. M. Lucantoni, "The BMAP/G/1 queue: A tutorial," in Models and techniques for Performance Evaluation of Computer and Communications Systems, L. D. a. R. Nelson, Ed., ed: Springer Verlag, 1993, pp. 330-58
5. D. M. Lucantoni, G. L. Choudhury, and W. Whitt, "The transient BMAP/G/1 queue," Stoch. Mod., vol. 10, pp. 145-82, 1994
6. D. M. Lucantoni, "Further transient analysis of the BMAP/G/1 queue," Stoch. Mod., vol. 14, pp. 461-78, 1998
Queueing Theory and Network Applications, Korea University, 2011
References (Cont’d)7. Heymen and Lucantoni, “Modeling Multiple IP Traffic Streams with Rate
Limits,” IEEE/ACM Trans. On Networking, 11, No. 6, 2003
8. Choudhury, Lucantoni and Whitt, “Multidimensional transform inversion with applications to the transient M/G/1 queue," Ann. Appl. Prob ., 4, No. 3, 719-740, 1994
7. L. Takács, Introduction to the Theory of Queues. New York: Oxford University Press, 1962
Queueing Theory and Network Applications, Korea University, 2011
Where to Find Additional References
Some Lucantoni papers can be downloaded from www.DLTconsulting.comFollow the links to my Publications page (or just
Google my name)More references are listed in those papers
Queueing Theory and Network Applications, Korea University, 2011