Queuing theory and traffic analysis in depth

Queuing Theory
and Traffic Analysis
CS 552
Richard Martin
Rutgers University

Queuing theory
• View network as collections of queues
– FIFO data-structures
• Queuing theory provides probabilistic
analysis of these queues
• Examples:
– Average length
– Probability queue is at a certain length
– Probability a packet will be lost

Little’s Law
• Little’s Law:
Mean number tasks in system = arrival rate x mean
response time
– Observed before, Little was first to prove
• Applies to any system in equilibrium,
as long as nothing in black box
is creating or destroying tasks
Arrivals Departures
System

Proving Little’s Law
J = Shaded area = 9
Same in all cases!
1 2 3 4 5 6 7 8
Packet
#
Time
1
2
3
1 2 3 4 5 6 7 8
# in
System
1
2
3
Time
1 2 3
Time in
System
Packet #
1
2
3
Arrivals
Departures

Definitions
• J: “Area” from previous slide
• N: Number of jobs (packets)
• T: Total time
 l: Average arrival rate
– N/T
• W: Average time job is in the system
– = J/N
• L: Average number of jobs in the system
– = J/T

1 2 3 4 5 6 7 8
# in
System
(L) 1
2
3
Proof: Method 1: Definition
Time (T)
1 2 3
Time in
System
(W)
Packet # (N)
1
2
3
=
W
L T
N
)
(

NW
TL
J 

W
L )
(l


Proof: Method 2: Substitution
W
L T
N
)
(

W
L )
(l

)
)(
( N
J
T
N
T
J

T
J
T
J
 Tautology

Example using Little’s law
• Observe 120 cars in front of the Lincoln
Tunnel
• Observe 32 cars/minute depart over a period where no
cars in the tunnel at the start or end (e.g. security
checks)
• What is average waiting time before and in
the tunnel?

Model Queuing System
Strategy:
Use Little’s law on both the complete system and its
parts to reason about average time in the queue
Server System
Queuing System
Queue Server
Queuing System
l
m
l

Kendal Notation
• Six parameters in shorthand
• First three typically used, unless specified
1. Arrival Distribution
• Probability of a new packet arrives in time t
2. Service Distribution
• Probability distribution packet is serviced in time t
3. Number of servers
4. Total Capacity (infinite if not specified)
5. Population Size (infinite)
6. Service Discipline (FCFS/FIFO)

Distributions
• M: Exponential
• D: Deterministic (e.g. fixed constant)
• Ek: Erlang with parameter k
• Hk: Hyperexponential with param. k
• G: General (anything)
• M/M/1 is the simplest ‘realistic’ queue

Kendal Notation Examples
• M/M/1:
– Exponential arrivals and service, 1 server, infinite
capacity and population, FCFS (FIFO)
• M/M/m
– Same, but M servers
• G/G/3/20/1500/SPF
– General arrival and service distributions, 3
servers, 17 queue slots (20-3), 1500 total jobs,
Shortest Packet First

M/M/1 queue model
l
m
l
m
1
Wq
W
L
Lq

Analysis of M/M/1 queue
• Goal: A closed form expression of the
probability of the number of jobs in the queue
(Pi) given only l and m

Solving queuing systems
• Given:
• l: Arrival rate of jobs (packets on input link)
• m: Service rate of the server (output link)
• Solve:
– L: average number in queuing system
– Lq average number in the queue
– W: average waiting time in whole system
– Wq average waiting time in the queue
• 4 unknown’s: need 4 equations

Solving queuing systems
• 4 unknowns: L, Lq W, Wq
• Relationships using Little’s law:
– L=lW
– Lq=lWq (steady-state argument)
– W = Wq + (1/m)
• If we know any 1, can find the others
• Finding L is hard or easy depending on the
type of system. In general:
0




n
n
nP
L

Equilibrium conditions
1
1
)
( 
 

 n
n
n P
P
P m
l
m
l
1:
2:
n+1
n
n-1
l l l
l
m m
m m
inflow = outflow
1 2
1

 n
n P
P m
l
3:
stability: 1
,
, 

 

m
l m
l

Solving for P0 and Pn
0
1 P
P 
   0
2
2 P
P 
   0
P
P
n
n 

1:




0
1
n
n
P 



0
0 1
n
n
P 

 
0
1
0
n
n
P

2:
3: 





0
1
,
1
1
n
n


 (geometric series)
,
, ,
, ,
4:











1
)
1
(
1
0
1
1
0
n
n
P 5:   )
1
( 
 

n
n
P

Solving for L
0




n
n
nP
L )
1
(
0





n
n
n 
 )
1
(
1
1






n
n
n


 



 

 1
1
)
1
( d
d






 

1
0
)
1
(
n
n
d
d


 
 
2
)
1
(
1
)
1
( 

 
 l
m
l



 
 )
1
(

Solving W, Wq and Lq
   l
m
l
l
m
l
l 
 

 1
1
L
W
    )
(
1
1
l
m
m
l
m
l
m
l
m 
 



W
Wq
)
(
)
(
2
l
m
m
l
l
m
m
l
l
l 
 

 q
q W
L

Response Time vs. Arrivals
l
m
 1
W
Waiting vs. Utilization
0
0.05
0.1
0.15
0.2
0.25
0 0.2 0.4 0.6 0.8 1 1.2
 %
W(sec)

Stable Region
Waiting vs. Utilization
0
0.005
0.01
0.015
0.02
0.025
0 0.2 0.4 0.6 0.8 1
 %
W(sec)
linear region

Empirical Example
M/M/m
system

Example
• Measurement of a network gateway:
– mean arrival rate (l): 125 Packets/s
– mean response time per packet: 2 ms
• Assuming exponential arrivals & departures:
– What is the service rate, m ?
– What is the gateway’s utilization?
– What is the probability of n packets in the
gateway?
– mean number of packets in the gateway?
– The number of buffers so P(overflow) is <10-6?

The service rate, m 
utilization =
P(n) packets in the gateway =
Example (cont)
pps
500
002
.
0
1

%
25
.
0
)
( 
 m
l

)
25
.
0
)(
75
.
0
(
)
)(
1
(
0
n
n
n
P
P 

 


Mean # in gateway (L) =
to limit loss probability to less than
1 in a million:
Example (cont)
33
.
0
25
.
0
1
25
.
0
1 
 


6
10

n


• Poisson process = exponential distribution
between arrivals/departures/service
• Key properties:
– memoryless
– Past state does not help predict next arrival
– Closed under:
– Addition
– Subtraction
Properties of a Poisson processes
t
e
t
P l



 1
)
arrival
(

Addition and Subtraction
• Merge:
– two poisson streams with arrival rates l1 and l2:
• new poisson stream: l3l1l2
• Split :
– If any given item has a probability P1 of “leaving”
the stream with rate l1:
 l2=(1-P1)l1

Queuing Networks
l1
l2
l3 l4
l5
4
2
6 l
l
l 

l6
l7
5
7 l
l 
0.3
0.5
0.7
0.5

Bridging Router Performance and
Queuing Theory
Sigmetrics 2004
Slides by N. Hohn*, D. Veitch*, K.
Papagiannaki, C. Diot

• End-to-end packet delay is an important
metric for performance and Service Level
Agreements (SLAs)
• Building block of end-to-end delay is through
router delay
• Measure the delays incurred by all packets
crossing a single router
Motivation

Overview
• Full Router Monitoring
• Delay Analysis and Modeling
• Delay Performance: Understanding and
Reporting

Measurement Environment
BackBone 1
BackBone 2
Customer 1

Packet matching
Set Link Matched pkts % traffic C2-out
C4 In 215987 0.03%
C1 In 70376 0.01%
BB1 In 345796622 47.00%
BB2 In 389153772 52.89%
C2 out 735236757 99.93%

Store & Forward Datapath
• Store: storage in input linecard’s
memory
• Forwarding decision
• Storage in dedicated Virtual Output
Queue (VOQ)
• Decomposition into fixed-size cells
• Transmission through switch fabric cell
by cell
• Packet reconstruction
• Forward: Output link scheduler
Not part of the system

Delays: 1 minute summary
MAX
MIN
Mean
BB1-In to C2-Out

memory
Queue (VOQ)
by cell
DliLj(L)

Minimum Transit Time
Packet size dependent minimum delay.

memory
Queue (VOQ)
by cell
DliLj(L)
FIFO queue

Modeling
Fluid queue with a delay element at the front

Modeling results
• A crude model performs well!
– As simpler/simpler than an M/M/1 queue
• Use effective link bandwidth
– account for encapsulation
• Small gap between router performance and queuing
theory!
• The model defines Busy Periods: time between the
arrival of a packet to the empty system and the time
when the system becomes empty again.

On the Delay Performance
• Model allows for router performance
evaluation when arrival patterns are known
• Goal: metrics that
– Capture operational-router performance
– Can answer performance questions directly
• Busy Period structures contain all delay
information
– BP better than utilization or delay reporting

Issues
• Report (A,D) measurements
• There are millions of busy periods even on a
lightly utilized router
• Interesting episodes are rare and last for a
very small amount of time

Duration of Congestion Level-L

Conclusions
• Results
– Full router empirical study
– Delay modeling
– Reporting performance metrics
• Future work
– Fine tune reporting scheme
– Empirical evidence of large deviations theory

Network Traffic Self-Similarity
Slides by Carey Williamson
Department of Computer Science
University of Saskatchewan

Introduction
• A classic measurement study has shown that
aggregate Ethernet LAN traffic is self-similar
[Leland et al 1993]
• A statistical property that is very different from
the traditional Poisson-based models
• This presentation: definition of network traffic
self-similarity, Bellcore Ethernet LAN data,
implications of self-similarity

Measurement Methodology
• Collected lengthy traces of Ethernet LAN
traffic on Ethernet LAN(s) at Bellcore
• High resolution time stamps
• Analyzed statistical properties of the resulting
time series data
• Each observation represents the number of
packets (or bytes) observed per time interval
(e.g., 10 4 8 12 7 2 0 5 17 9 8 8 2...)

• If you plot the number of packets observed
per time interval as a function of time, then
the plot looks ‘‘the same’’ regardless of what
interval size you choose
• E.g., 10 msec, 100 msec, 1 sec, 10 sec,...
• Same applies if you plot number of bytes
observed per interval of time
Self-Similarity: The intuition

• In other words, self-similarity implies a
‘‘fractal-like’’ behavior: no matter what time
scale you use to examine the data, you see
similar patterns
• Implications:
– Burstiness exists across many time scales
– No natural length of a burst
– Key: Traffic does not necessarily get ‘‘smoother”
when you aggregate it (unlike Poisson traffic)
Self-Similarity: The Intuition

Self-Similarity Traffic Intuition (I)

Self-Similarity in Traffic Measurement II

• Self-similarity is a rigorous statistical property
– (i.e., a lot more to it than just the pretty ‘‘fractal-
like’’ pictures)
• Assumes you have time series data with finite
mean and variance
– i.e., covariance stationary stochastic process
• Must be a very long time series
– infinite is best!
• Can test for presence of self-similarity
Self-Similarity: The Math

• Self-similarity manifests itself in several
equivalent fashions:
• Slowly decaying variance
• Long range dependence
• Non-degenerate autocorrelations
• Hurst effect
Self-Similarity: The Math

Methods of showing Self-Similarity
H=0.5
H=0.5
H=1
Estimate H  0.8

• The variance of the sample decreases more
slowly than the reciprocal of the sample size
• For most processes, the variance of a sample
diminishes quite rapidly as the sample size is
increased, and stabilizes soon
• For self-similar processes, the variance
decreases very slowly, even when the sample
size grows quite large
Slowly Decaying Variance

• The ‘‘variance-time plot” is one means to test
for the slowly decaying variance property
• Plots the variance of the sample versus the
sample size, on a log-log plot
• For most processes, the result is a straight
line with slope -1
• For self-similar, the line is much flatter
Time-Variance Plot

Variance-Time Plot
Variance
m
Variance of sample
on a logarithmic scale
0.0001
0.001
10.0
0.01
100.0

Variance-Time Plot
Variance
m
Sample size m
1 10 100 10 10 10 10
4 5 6 7

Variance-Time Plot
Variance
m
Slope = -1
for most processes

Variance-Time Plot
Variance
m
Slope flatter than -1
for self-similar process

• Correlation is a statistical measure of the
relationship, if any, between two random
variables
• Positive correlation: both behave similarly
• Negative correlation: behave as opposites
• No correlation: behavior of one is unrelated to
behavior of other
Long Range Dependence

• Autocorrelation is a statistical measure of the
relationship, if any, between a random
variable and itself, at different time lags
• Positive correlation: big observation usually
followed by another big, or small by small
• Negative correlation: big observation usually
followed by small, or small by big
• No correlation: observations unrelated

• Autocorrelation coefficient can range
between:
+1 (very high positive correlation)
-1 (very high negative correlation)
• Zero means no correlation
• Autocorrelation function shows the value of
the autocorrelation coefficient for different
time lags k

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Autocorrelation Function

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Maximum possible
positive correlation

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Maximum possible
negative correlation

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
No observed
correlation at all

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Significant positive
correlation at short lags

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
No statistically significant
correlation beyond this lag

• For most processes (e.g., Poisson, or
compound Poisson), the autocorrelation
function drops to zero very quickly
– usually immediately, or exponentially fast
• For self-similar processes, the autocorrelation
function drops very slowly
– i.e., hyperbolically, toward zero, but may never
reach zero
• Non-summable autocorrelation function

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Typical short-range
dependent process

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Typical long-range
dependent process

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Typical long-range
dependent process
Typical short-range
dependent process

• For self-similar processes, the autocorrelation
function for the aggregated process is
indistinguishable from that of the original
process
• If autocorrelation coefficients match for all
lags k, then called exactly self-similar
• If autocorrelation coefficients match only for
large lags k, then called asymptotically self-
similar
Non-Degenerate Autocorrelations

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Original self-similar
process

+1
-1
0
lag k
0 100
Autocorrelation
Coefficient
Original self-similar
process
Aggregated self-similar
process

• Aggregation of a time series X(t) means
smoothing the time series by averaging the
observations over non-overlapping blocks of
size m to get a new time series Xm(t)
Aggregation

• Suppose the original time series X(t) contains
the following (made up) values
2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
• Then the aggregated series for m = 2 is:
Aggregation Example

the following (made up) values:
2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
Aggregation Example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
4.5
Aggregation Example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
4.5 8.0
Aggregation example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
4.5 8.0 2.5
Aggregation Example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
4.5 8.0 2.5 5.0
Aggregation Example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
4.5 8.0 2.5 5.0 6.0 7.5 7.0 4.0 4.5 5.0...
Aggregation Example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
Then the aggregated time series for m = 5 is:
Aggregation Example

Aggregation: An Example
2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
6.0

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
6.0 4.4

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
6.0 4.4 6.4 4.8 ...

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
5.2

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1...
5.2 5.6

• For almost all naturally occurring time series,
the rescaled adjusted range statistic (also
called the R/S statistic) for sample size n
obeys the relationship
E[R(n)/S(n)] = c nH
where:
R(n) = max(0, W1, ... Wn) - min(0, W1, ... Wn)
S2(n) is the sample variance, and
for k = 1, 2, ... n
Hurst Effect

• For models with only short range
dependence, H is almost always 0.5
• For self-similar processes, 0.5 < H < 1.0
• This discrepancy is called the Hurst Effect,
and H is called the Hurst parameter
• Single parameter to characterize self-similar
processes
Hurst Effect

R/S Statistic: An Example
2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
• There are 20 data points in this example

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
• For R/S analysis with n = 1, you get 20
samples, each of size 1:

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 1: X = 2, W = 0, R(n) = 0, S(n) = 0
n 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 2: X = 7, W = 0, R(n) = 0, S(n) = 0
n 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 1: X = 4.5, W = -2.5, W = 0,
R(n) = 0 - (-2.5) = 2.5, S(n) = 2.5,
R(n)/S(n) = 1.0
n 1 2

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 2: X = 8.0, W = -4.0, W = 0,
R(n) = 0 - (-4.0) = 4.0, S(n) = 4.0,
R(n)/S(n) = 1.0
n 1 2

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 1: X = 4.3, W = -2.3, W = 0.3, W = 0
R(n) = 0.3 - (-2.3) = 2.6, S(n) = 2.05,
R(n)/S(n) = 1.30
n 1 2 3

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 2: X = 5.7, W = 6.3, W = 5.7, W = 0
R(n) = 6.3 - (0) = 6.3, S(n) = 4.92,
R(n)/S(n) = 1.28
n 1 2 3

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 1: X = 6.0, W = -4.0, W = -3.0,
W = -5.0 , W = 1.0 , W = 0, S(n) = 3.41,
R(n) = 1.0 - (-5.0) = 6.0, R(n)/S(n) = 1.76
n 1 2
3 4 5

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
Block 2: X = 4.4, W = -4.4, W = -0.8,
W = -3.2 , W = 0.4 , W = 0, S(n) = 3.2,
R(n) = 0.4 - (-4.4) = 4.8, R(n)/S(n) = 1.5
n 1 2
3 4 5

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1

2 7 4 12 5 0 8 2 8 4 6 9 11 3 3 5 7 2 9 1
sample of size 20:

• Another way of testing for self-similarity, and
estimating the Hurst parameter
• Plot the R/S statistic for different values of n,
with a log scale on each axis
• If time series is self-similar, the resulting plot
will have a straight line shape with a slope H
that is greater than 0.5
• Called an R/S plot, or R/S pox diagram
R/S Plot

R/S
Statistic
Block Size n
R/S Pox Diagram

R/S
Statistic
Block Size n
R/S statistic R(n)/S(n)
R/S Pox Diagram

R/S
Statistic
Block Size n
Sample size n
R/S Pox Diagram

R/S
Statistic
Block Size n
Slope 0.5
R/S Pox Diagram

R/S
Statistic
Block Size n
Slope 0.5
Slope 1.0
R/S Pox Diagram

R/S
Statistic
Block Size n
Slope 0.5
Slope 1.0
Self-
similar
process
R/S Pox Diagram

R/S
Statistic
Block Size n
Slope 0.5
Slope 1.0
Slope H
(0.5 < H < 1.0)
(Hurst parameter)
R/S Pox Diagram

• Self-similarity is an important mathematical
property that has recently been identified as
present in network traffic measurements
• Important property: burstiness across many
time scales, traffic does not aggregate well
• There exist several mathematical methods to
test for the presence of self-similarity, and to
estimate the Hurst parameter H
• There exist models for self-similar traffic
Self-Similarity Summary

Newer Results
V. Paxson, S. Floyd, Wide-Area Traffic: The Failure of
Poisson Modeling, IEEE/ACM Transaction on Networking,
1995.
• TCP session arrivals are well modeled by a Poisson
process
• A number of WAN characteristics were well modeled by
heavy tailed distributions
• Packet arrival process for two typical applications (TELNET,
FTP) as well as aggregate traffic is self-similar

Another Study
M. Crovella, A. Bestavros, Self-Similarity in World Wide
Web Traffic: Evidence and Possible Causes,
IEEE/ACM Transactions on Networking, 1997
• Analyzed WWW logs collected at clients over a 1.5
month period
– First WWW client study
– Instrumented MOSAIC
• ~600 students
• ~130K files transferred
• ~2.7GB data transferred

Self-Similar Aspects of Web traffic
• One difficulty in the analysis was finding stationary,
busy periods
– A number of candidate hours were found
• All four tests for self-similarity were employed
– 0.7 < H < 0.8

Explaining Self-Similarity
• Consider a set of processes which are either
ON or OFF
– The distribution of ON and OFF times are heavy
tailed
– The aggregation of these processes leads to a
self-similar process
• So, how do we get heavy tailed ON or OFF
times?

• Analysis of client logs showed that ON times were, in fact,
heavy tailed
– Over about 3 orders of magnitude
• This lead to the analysis of underlying file sizes
– Over about 4 orders of magnitude
– Similar to FTP traffic
• Files available from UNIX file systems are typically heavy tailed
Impact of File Sizes

Heavy Tailed OFF times
• Analysis of OFF times showed that they are
also heavy tailed
• Distinction between Active and Passive OFF
times
– Inter vs. Intra click OFF times
• Thus, ON times are more likely to be cause of
self-similarity

Major Results from CB97
• Established that WWW traffic was self-similar
• Modeled a number of different WWW
characteristics (focus on the tail)
• Provide an explanation for self-similarity of WWW
traffic based on underlying file size distribution

Where are we now?
• There is no mechanistic model for Internet traffic
– Topology?
– Routing?
• People want to blame the protocols for observed behavior
• Multiresolution analysis may provide a means for better
models
• Many people (vendors) chose to ignore self-similarity
– Does it matter????
– Critical opportunity for answering this question.

Queuing theory and traffic analysis in depth

Recommended

More Related Content

Similar to Queuing theory and traffic analysis in depth (20)

Recently uploaded (20)

Queuing theory and traffic analysis in depth