PPKE ITK
2011/12tanév
Őszifélév
Tájékoztatáshttp://digitus.itk.ppke.hu/~gosztony/
3.6. Multi-service queueing systems
3.7. Queueing networks
4. Traffic measurement
Final overview
Infocomm network’s planningtraffic aspects
2Infocomm network’s planning - traffic aspects - 2011.10.05.
3.6
Textbook: Chapter 11
Multi-service queueing systems
3Infocomm network’s planning - traffic aspects - 2011.10.05.3
Multi-service queueing systems – 1.There are more than one type (class, service, stream) of customers
Customers of same type (class, service, stream) belong to a specfic chain, a queueing system is a node in a queueing network
In reversible systems , the departure process is of the same type as the arrival process, in the investigated cases: Poisson processes. In this case several queueing systems may be combined into a network of queueing systems.
In classical queueing systems we have non-sharing. A customer is either waiting or being served. When being served it has a server alone. There are queueing systems with different sharing strategies. Customers may share servers with other customers so that no one is waiting, but always served with some rate, which may be smaller than the requested rate.
By requesting reversibility and usage of all servers whenever possible we get the processor sharing (PS) and generalised processor sharing strategy (GPS).
4Infocomm network’s planning - traffic aspects - 2011.10.05.4
Multi-service queueing systems – 2.
In multi-service queueing systems and in queueing networks customers in some way share the available capacity, and therefore they are served all the time. But they may obtain less capacity than requested, which results in an increase of sojourn time. The sojourn time is not split up into separate waiting time and service time. For multiservice queueing systems and for queueing networks the definitions below are used:
5Infocomm network’s planning - traffic aspects - 2011.10.05.5
Multi-service queueing systems – 2.
jjj sWW
In multi-service queueing systems and in queueing networks the waiting time is defined as the total sojourn time , including the service time, s. E.g. if in the Internet, the bandwidth is smaller than requested, then the mean transfer time Wj will be bigger then sj and the increase is defined as the mean virtual waiting time:
In a similar way the mean virtual queue length i.e. Lj is defined as:
jjj ALL
where Aj is the offered traffic of type j.
6Infocomm network’s planning - traffic aspects - 2011.10.05.6
Multi-service queueing systems – 3.Example.
In general there are N PCT1 input processes with j arrival and μ1 departure intensities (j = 1 … N)
(The exact value of the μi departure intensities requires further consideration !)
7Infocomm network’s planning - traffic aspects - 2011.10.05.7
Multi-service queueing systems – 4.
PS: Processor SharingPR: Preemptive Resume
8Infocomm network’s planning - traffic aspects - 2011.10.05.
3.7
Queueing networks
Introduction to queueing networks Symmetric queueing systems, Jackson’s theorem Closed networks: single chain Closed networks: several chains Other questions
Textbook: Chapter 12
9Infocomm network’s planning - traffic aspects - 2011.10.05.
A queueing system consists of nodes (group of servers with the same function) and jobs (requests) travelling from node to node.
In queueing networks we define the queue-length in a node as the total number of jobs in the node, including delayed and served jobs.
A queueing network might be:1. open – the number of jobs may change, i. e. M/M/n
2. closed– the number of jobs is fixed, i.e. Palm’s machine repair model
The departure process from one node is the arrival process at another node, special attention should be payed to the departure process
Queueing networks, introduction – 1.
10Infocomm network’s planning - traffic aspects - 2011.10.05.
Queueing networks, introduction – 2.
Four nodes.Four open chains.
State:
where: gives the number of jobs in node k
and
pj,k is the probability, that a jobt having left node j. is directed to node k.
p22
11Infocomm network’s planning - traffic aspects - 2011.10.05.
The queueing system is symmetric if both the arrival and the departure processes are Poissonian.
The four models:• M/M/n
state probabilities
and
• M/G/∞*state probabilities(Poisson !)
• M/G/1–PS* state probabilities
• M/G/1-LCFS-PR* state probabilities * immediate service!
Symmetric systemsR
evers
ibili
ty
PS = Processor SharingPR = Preemptive Resume
12Infocomm network’s planning - traffic aspects - 2011.10.05.
Jackson’s theorem– 1a. Jackson’s theorem: Consider an open queueing network with K nodes satisfying the following conditions:
13Infocomm network’s planning - traffic aspects - 2011.10.05.
Jackson’s theorem– 1b.
14Infocomm network’s planning - traffic aspects - 2011.10.05.
Jackson’s theorem – 1c.
The key point of Jackson's theorem is that:
each node can be considered independently of all other nodes and that the state probabilities are as for Erlang's delay system. (Erlang’s C formula.) This simplifies the calculation of the state space probabilities significantly.
Jackson's first model thus only deals with open queueing networks.
Example:
Open queuing network consisting of two M|M|1 systems in series
15Infocomm network’s planning - traffic aspects - 2011.10.05.
Jackson’s theorem – 2.
In Jackson's second model the arrival intensity from outside:
may depend on the current number of customers in the network,
furthermore, μk can depend on the number of customers at node k.
In this way, we can model queueing networks whichare either closed, open, or mixed. In all three cases, the state probabilities have product form.
16Infocomm network’s planning - traffic aspects - 2011.10.05.
Independence assumption
Kleinrock’s independence assumption
If we consider a real-life data network, then the packets will have the same constant length, and therefore the same service time on all links and nodes of equal speed. The theory of queueing networks assumes that a packet (a customer) samples a new service time in every node. This is a necessary assumption for the product form. This assumption was first investigated by Kleinrock (1964 [74]), and it turns out to be a good approximation in praxis.
17Infocomm network’s planning - traffic aspects - 2011.10.05.
Single open chain
Open system
One has to find the state probabilities
where ik is the number of requests in node k.
Steps: 1. solution of
2. using μi –s one may get Ai –k.
3. by considering Erlang's delay system one may get the state probabilities for each node.
18Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 1. Convolution algorithm for closed networksWe only know the relative load at each node, not the absolute load, i.e. cΛj is obtained, but c is unknown. We can obtain the relative state probabilities. Finally, by normalizing we get the normalized state probabilities. Large system complex run.Steps:
19Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 2. Convolution algorithm for closed networks
20Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 3.
21Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 4. Terminals M/G/1 – IS* node
CPU M/M/1 node
*IS = Immediate Service – There is always a free terminal for the new task.
Example 12.5.1
λ1= λ
λ2= λ
See details inthe Textbook
22Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 5. Assumptions: S constant.
S circulating jobs. The CPU andthe I/O channelsserve each job several times. A departing job willimmediatly replaced by a new on.
Example 12.5.2 S = 4K = 3 (CPU + 2 I/O)
exponentialholding timess = 1/μ
See details inthe Textbook
23Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 6.
MVA (Mean Value Algorithm)
K nodes, S jobs (in a single chain), αk = λk sk relative traffics.Recursion according the number x of jobs.At node k there are Lk(S) jobs on the average.
Rem
em
ber!
24Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 7.
PS = Processor SharingPR = Preemptive Resume
average sojourn time
immediate service
25Infocomm network’s planning - traffic aspects - 2011.10.05.
There is no waiting in the system, if S=1
Single closed-chain – 8.
Example for nk = 1 but the method might be generalized for any nk .
kk sW 1 kk sW 1
26Infocomm network’s planning - traffic aspects - 2011.10.05.
Single closed-chain – 9. Example 12.5.3 (= 12.5.2 but with MVA)
K=3, S=4,
relative k values:
Recursion formulae:i
iiii s
aa
ii
1s
See details inthe Textbook
27Infocomm network’s planning - traffic aspects - 2011.10.05.
BCMP queueing networks
Queueing networks with more than one type ofcustomers also have product form state probabilities.(Generalization of Jackson’s second model. BCMP Baskett,Chandy, Muntz és Palacios - 1975)
Necessary conditions:
BCMP–networks can be evaluated with the multi-dimensional convolution algorithm and the multidimensional MVA algorithm.
28Infocomm network’s planning - traffic aspects - 2011.10.05.
Mixed queueing networks
Mixed queueing networks (open & closed) are calculated by first calculating the traffic load in each node from the open chains. This traffic must be carried to enter statistical equilibrium. The capacity of the nodes are reduced by this traffic, and the closed queueing network is calculated by the reduced capacity. So the main problem is to calculate closed networks. For this we have more algorithms among which the most important ones are convolution algorithm and the MVA (Mean Value Algorithm) algorithm.
29Infocomm network’s planning - traffic aspects - 2011.10.05.
Complexity
30Infocomm network’s planning - traffic aspects - 2011.10.05.
4.
Traffic measurement
Principles and methods Theory of sampling Continuous measurement Scanning
Textbook: Chapter 13
31Infocomm network’s planning - traffic aspects - 2011.10.05.
Introduction Traffic measurement gathering of data about the
traffic of real or fictisous systems which serve requests arriving in a random way.
A minimum of technical and administrative efforts should result in a maximum of information and benefit.
A measurement during a limited time intervalcorresponds to a registration of a certain realization of the traffic process.
Margin of error condfidence interval.
For practical purposes it is in general sufficientto know the mean value and the variance.
32Infocomm network’s planning - traffic aspects - 2011.10.05.
Principles, methods – 1.
What do we measure? How do we measure?
Eventsarrival of requests, number of jobs,number of lost requests,etc.
Time intervalsholding times,waiting times,execution times of jobs,etc.
Continuous measurementthe measuring point is active, it activates the measuring equipment at the instant of the event
Scanningthe measuring point is passive,the measuring equipment itself tests from time to time, if any changes have taken place.
33Infocomm network’s planning - traffic aspects - 2011.10.05.
Principles, methods – 2.
Continuous measurement - examples:
• electro-mechanical counters (e.g. number of copies)
• x-y plotters (earthquake survaillance)• Ampere-hour meters• water meters
Scanning - examples:
• call-charging impulses• measuring of carried traffic by repeated scans
34Infocomm network’s planning - traffic aspects - 2011.10.05.
Principles, methods – 3. Busy state of servers:
actual
scanned
35Infocomm network’s planning - traffic aspects - 2011.10.05.
Principles, methods – 4.
Registration of changes
36Infocomm network’s planning - traffic aspects - 2011.10.05.
Theory of sampling – 1.
A sample of n IDD* observations withunknown finite mean value m1 and finite varianceб2 are available (population parameters)
The average value and variance of the sample:
These are functions of arv., they are rv.-s also (sample distribution) andrepresent an estimationof the average value and variance of the unknown population:
(correctedempiricalvariance)
*IDD= Independent and Identically Distributed
37Infocomm network’s planning - traffic aspects - 2011.10.05.
Theory of sampling – 2.
The accuracy of an estimate of a sample parameter is described by means of a confidence interval :
This is valid, if the samples are independent. Independence is fulfilled, if measurements take place on different days, but samples are not independent for samples taken by scanning in a limited time interval.
38Infocomm network’s planning - traffic aspects - 2011.10.05.
Theory of sampling – 3.Percentiles of the t−distribution with n degrees of freedom. A specific value of α corresponds to a probability mass α/2 in bothtails of the t−distribution. When n is large, then we may use the percentiles of the Normal distribution.
39Infocomm network’s planning - traffic aspects - 2011.10.05.
Theory of sampling – 4.
Mean valueof sample
Correctedempiricalvariance
Example 13.2.1:
40Infocomm network’s planning - traffic aspects - 2011.10.05.
Continuous measurement– 1.
Relationships valid forstochastic sums might in principle be applied if the measuring periodis unlimited.
Application in practiceis possible withprecaution !
Time intervals markedby full lines are measured in casea. and b., respectively.
41Infocomm network’s planning - traffic aspects - 2011.10.05.
Stochastic sum - 1.From the mathematical point of view the evaluation of the measurement consist of calculating the sum of a random amount of random variables
Service without congestion.Arrival process and holding times are independent.
The number of arriving requests in the given period of length T is a random variable: N.
Ti. gives the holding time of thei-th incoming request. Distribution of Ti –s is uniform.
Total traffic in T
See 2.3.3 of Textbook
42Infocomm network’s planning - traffic aspects - 2011.10.05.
Stochastic sum - 2.
Ti and N are stochastically independent
A stochastic sum may be interpreted as a series/parallel combination of random variable.
The problem in graphical form:
43Infocomm network’s planning - traffic aspects - 2011.10.05.
Stochastic sum 3.For a given i-th branch:
For all i branches:
variance, since the numberof holding times is an r.v.
variance, since the holdingtime is an r.v.
mean value
variance
second non-central moment
44Infocomm network’s planning - traffic aspects - 2011.10.05.
Continuous measurement – 2.Applying the relationships of the „stochastic sum”:
The amount of traffic (holding times) and trafficintensity (arrival of requests) should be independent.This is fulfilled, if congestion is negligible.Assumption: Poisson input process
Number of requests in interval T:
Thus the requiredservice time i.e. traffic:
The first and second non-central momentsof the holding time.
Palm’s form factor
45Infocomm network’s planning - traffic aspects - 2011.10.05.
Continuous measurement – 3.
The distribution ST of the stochastic sum is a compound Poisson distribution.
represents amount of traffic.
The average value of the number occupied serversrepresents traffic intensity = the amount of trafficper time unit if the time unit equals theaverage holding time:
Valid for any holding time distribution !
46Infocomm network’s planning - traffic aspects - 2011.10.05.
Continuous measurement – 4.
Independent of holding time distribution.
Depends on holding time distribution
Relative accuracy of measurement:
The measurement of smaller intensities is more precise!!
2.9 )
See the followingFigure !
47Infocomm network’s planning - traffic aspects - 2011.10.05.
Figure 1.18: Frequency function for holding times of trunks in a local switching centre.
Result of holding time measurements
48Infocomm network’s planning - traffic aspects - 2011.10.05.
Scanning measurement – 1.
Constant scanning interval (h)
The continuously distributed holding timeis approximated witha discretely distributed holding time.
The continuous timeintervals may overlap.Estimation is thereforemore difficult.
49Infocomm network’s planning - traffic aspects - 2011.10.05.
If the real holding times have a distribution function F(t), then it can be shown that one will observe the following discrete distribution :
It can also be shown that for any distribution function F(t) one will always observe the correct mean value:
Scanning measurement – 2.
50Infocomm network’s planning - traffic aspects - 2011.10.05.
For exponential holding times –the so called Westerberg distribution may be observed:
Scanning measurement – 3.
!!!
51Infocomm network’s planning - traffic aspects - 2011.10.05.
Sampling measurement
2A/T ≠
but it =, if h 0
Measurement by scanning = sampling from a sample!!
Scanning measurement – 4.
Continuous measurement
for exponential distribution:
52Infocomm network’s planning - traffic aspects - 2011.10.05.
Final overview – 1. Info-communication systems handle service requests
the totality of which is simply denominated as traffic. Traffic handling capacity has an influence on service
quality. Arrivals and holding times of service requests in info-
communication systems have random nature, their forecasting is not simple.
Traffic handling capacity can be estimated by mathematical models and computer simulation.
Nowadays the importance of network traffic management methods e.g. software-based routing algorithms is increasing.
Network planning requires some understanding of traffic engineering.
53Infocomm network’s planning - traffic aspects - 2011.10.05.
Final overview –2.
TTE tasks planning,
performance evaluation, operation,
maintenanceof infocomm. systems
infocom trends
rapidly growingtraffic demands
What is traffic ?How does it look like ?
intensity ofincomingrequests,
holding times
GoS QoSSLA (Service Level Agreement)Network Traffic Management
54Infocomm network’s planning - traffic aspects - 2011.10.05.
Final overview –3.
What happens to rejected traffic
requests ?
loss, rerouting,limited waiting,
unlimited waiting
Erlang’loss model,constant intensity of requests
Engset’s loss model,state dependent intensity of requests
BPPtraffic models
(PCT)
Overflow traffic has a peaked character !!
Network traffic
management
Best service with available resources in a randomly fluctuating situation
55Infocomm network’s planning - traffic aspects - 2011.10.05.
Final overview –4.
Specification Strategies Priorities Behaviour of requests
Queuing systems
M|M|nErlang
+ M|M|1|S|SPalm
Generalresults
Little’s theoremPollaczek-Khintchine’s formula
Systems mentioned:
M|G|1, M|G|1|kM|D|1, M|D|n
GI|G|1(some with priorities)
processor sharing
56Infocomm network’s planning - traffic aspects - 2011.10.05.
Final overview –5.
Queueing with shared serviceMean virtual waiting timemean virtual queue length
Several chains of jobsreversibility
links and nodes structure
Multiservicequeueing systems
Queueing networks
ClosedOpen
Convolution algorithmMean Value AlgorithmJackson’s theorems:
1. 2.
Kleinrock’sindependence assumption !