infocom 2013-2-state-markov
DESCRIPTION
TRANSCRIPT
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Variants of 2-state Markov Models – Gilbert-Elliott Channels– Semi-Markov Processes SMP(2)
Formula for the 2nd Order Statistics of 2-State Models Model Adaptation to Traffic Profiles Conclusions and Outlook
2-state (semi-)Markov Processes beyond Gilbert-Elliott:Traffic and Channel Models based 2nd Order Statistics
Gerhard Haßlinger1, Anne Schwahn2, Franz Hartleb2
1Deutsche Telekom Technik, 2T-Systems, Darmstadt, Germany
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
GoodState
BadState
q
Gilbert-Elliott channel: 4 parameters (p,q,hG,hB)
1 – p1 – qp
GoodState
BadState
q, hGB
2-state Markov process with transition specific rates: 6 parameters (p,q,hGG,hGB,hBG,hBB)
1 – phBB
1 – qhGG p, hBG
State G q
2-state semi-Markov process for traffic rate distributions RG( );RB( )6 param. (p,q,G,G
2,B,B2) in 2nd order statistics
1 – p1 – qp
hG hB
RG( );G;G
2
State BRB( );B;B
2
GoodState
BadState
q
Gilbert-Elliott channel: 4 parameters (p,q,hG,hB)
1 – p1 – qp
GoodState
BadState
q, hGB
2-state Markov process with transition specific rates: 6 parameters (p,q,hGG,hGB,hBG,hBB)
1 – phBB
1 – qhGG p, hBG
State G q
2-state semi-Markov process for traffic rate distributions RG( );RB( )6 param. (p,q,G,G
2,B,B2) in 2nd order statistics
1 – p1 – qp
hG hB
RG( );G;G
2
State BRB( );B;B
2
2-State (semi-)Markov Processes
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Application spectrum of 2-state Markov models
Traffic profiles, dimensioning for QoS/QoE demands- many papers on measurement of traffic profiles- many papers on queueing analysis with 2-(M-)state Markov input
Error channel modeling- many papers on channel profiles (e.g., Rician fading, etc.)- some papers on error models for packets, data blocks of protocols- many papers on performance of error-detecting/correcting codes
Application examples in other disciplines- in economics: for volatility in markets- in nuclear physics: for electron spin signals- in statistics of medicine: for estimation of misclassification- in documentation: for modeling of image degradation- analytical verification of simulations etc.
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Measured 2nd order statistics over several time scales =1 10 100 1000
0,0
0,5
1,0
1,5
2,0
0,001 0,01 0,1 1 10 100 1000Time Scale [s]
Stan
dard
Dev
iatio
n / M
ean
Rat
e
TwitterFacebookUploadedVoIPYouTubeTotal Traffic
=1 10 100 1000
0,0
0,5
1,0
1,5
2,0
0,001 0,01 0,1 1 10 100 1000Time Scale [s]
Stan
dard
Dev
iatio
n / M
ean
Rat
e
TwitterFacebookUploadedVoIPYouTubeTotal Traffic
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
)(
)1(11)1(212
2
qpNqp
qpqp
N
N
N
Results for the 2nd order statistics
2. 2-state Markov with
;2222 HN N1. Self-similar traffic:
Adaptation to traffic profile with mean rate and variance on smallest measurement time scale (1ms time slots): G , B are determined , only one parameter p+q remains free in the 2nd order statistics
Remark: 2nd order statistics is equivalent to autocorrelation function
3 Parameters: , , H
H: Hurst Parameter (0.5 < H < 1)
4 Parameters: p, q, G , B
constant rate in each state:
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Results for the 2nd order statistics
)(
)1(11)(11 2222
qpNqp
qpqp
N
N
N
3. Markov modulated Poisson process MMPP(2) ( 22):
4. Semi-Markov process SMP(2):
;)(
)(2
;)()1(111
][
)(
2
22 ][
GBBGBGBG
N
N
qpEE
qpEEpq
qpNqp
N
.)1(;)1(
BGBBB
GBGGG
ppEqqE
4 Parameters: p, q, G , B (G2=G
2, B2=B
2); only one parameter p+q remains free in the 2nd order statistics
6 Parameters: p, q, G , B , G , B ; or 10 param.: p, q, GG , GB , GB , BB , GG , GB , GB , BB
2 parameters , p+q remain free in the 2nd order statistics
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
SMP(2) Fitting of 2nd Order Statistics
0
5
10
15
20
25
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192s
Time scale
Stan
dard
Dev
iatio
n [M
b/s]
2-state SMP (p=5q=0.00001)2-state SMP (p=5q=0.0005)2-state SMP (p=5q=0.0028)2-state SMP (p=5q=0.05)Measurement Result2-sate SMP (p=5q=5/6)
1. Step of parameter fitting: p/q = 5 is const.; p+q is variable Monotonous increase of 2
8192 for p+q 0; match at p = 5q = 0.0028
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
SMP(2) Fitting of 2nd Order Statistics
2. Step of parameter fitting: p/q is variable; 28192 is kept constant;
Monotonous decrease of N=013 2
2N ; best match for p/q = 0.0013
0
50
100
150
200
250
300
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192s
Time scale
Stan
dard
dev
iatio
n [M
b/s]
SMP(2) with p/q = 0.405 (max.)SMP(2) with p/q = 0.1SMP(2) with p/q = 0.04Measurement ResultSMP(2) with p/q = 0.013SMP(2) with p/q = 0.00923 (min.)
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Fitting of the 2nd order statistics for YouTube traffic
0
40
80
120
160
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192sTime scale
Stan
dard
dev
iatio
n[M
b/s]
Fixed Rate per StateSelf-Similar ProcessMMPP(2)Measurement ResultSMP(2)
All models are fitted to µ, 12 and 2
8192 ; A least mean square deviation criterion could be fitted in a 3. step, which isn´t monotonous optimization
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Fitting of the 2nd order statistics for Facebook traffic
0
5
10
15
20
25
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192s
Time scale
Stan
dard
Dev
iatio
n [M
b/s]
Fixed Rate per StateMMPP(2)Self-Similar ProcessMeasurement ResultSMP(2)
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Fitting of the 2nd order statistics for RapidShare traffic
0
5
10
15
20
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192sTime scale
Stan
dard
dev
iatio
n [M
b/s]
Fixed Rate per StateSelf-Similar ProcessMMPP(2)Measurement ResultSMP(2)
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Fitting of the 2nd order statistics for the total traffic
MMPP(2) fitting curve is missing, since 12 < µ2 cannot be achieved
0
50
100
150
200
250
300
0.001s 0.004s 0.016s 0.064s 0.256s 1.024s 8.192s
Time scale
Stan
dard
dev
iatio
n [M
b/s]
Fixed Rate per State
Self-Similar Process
Measurement Result
SMP(2)
2-State (Semi-)Markov Models & 2nd Order Statistics
GerhardHasslinger
Turin April, 17th 2013
Conclusions on 2-state traffic models Explicit formula for the 2nd order statistics of 2-state (semi-)Markov
SMP(2) processes clearly reveals impact of parameters- More complex Eigenvalue solutions for N-state Markov
SMP(2) model variants with 6 parametersprovide a 2-dimensional adaptation space (p, q) fairly good fitting of measured traffic variability in times scales
from 1ms to 10s
Gilbert-Elliott, MMPP(2) and self-similar models have onlyone parameter for 2nd order adaptation only coarse fitting accuracy for measured traffic variability
Traffic models of superposed or otherwise combined 2-state models have potential for improvement