You are on page 1of 32

OR

OPERATIONS RESEARCH

QUEUEING THEORY & MODELS

Course facilitators: Prof. Ajith Kumar Prof. T N Badri


Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING THEORY
The motivation to study queueing systems
FROM THE PERSPECTIVE OF A COUNTRY

OR

In the US alone, people collectively spend 37,000,000,000 manhours per year waiting in queues (Hillier and Lieberman, 2005) This amounts to an opportunity loss of nearly 1.5 billion man-days of work per year! (this is for the US alone).

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING THEORY
The motivation to study queueing systems
FROM THE PERSPECTIVE OF A COMPANY

OR

Increasing service counters decreases customer waiting time and increases customer satisfaction. But costs of providing service goes up. Decreasing service counters increases customer waiting time and decreases customer satisfaction. It decreases cost of providing service, but cost of customer dissatisfaction goes up. Total cost = Cost of providing service + Cost of customer dissatisfaction If you reduce one of the components, the other one increases. SO, WHAT IS THE OPTIMAL SERVICE LEVEL?
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING THEORY
The basic queueing model

OR

Input Characteristics

MODEL OF A QUEUEING SYSTEM

Output Characteristics

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING THEORY
The inside of a queueing system
customer does not rejoin population customer rejoins the population

OR
Calling population

server

Waiting line (Q)

S
customer receiving service (Not in the Q) customers waiting (in the Q)

A single-channel, single-phase system


Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


S

OR
Single channel, Single phase Single channel, Multi phase

S3

S2

S1

S1 S2 S3 T1 U T2 T3 S1 S2 S3
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

Multi channel, Single phase

Multi channel, Multi phase

QUEUEING SYSTEMS INPUTS


Inputs These are the typical inputs considered Arrival rate, arrival time, and arrival size. Service characteristics (service rate, service time). Capacities (of the systems, of the sub-systems). Type of population Q discipline (FIFO, LIFO, SIRO, SPT, PR) Q behavior (balking, reneging, jockeying) Exit fates (rejoins population, does not rejoin).

OR

Each of these is explored briefly in the next few slides

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Arrival Characteristics
ARRIVAL RATE
No. of customers that arrive in a given period of time.

OR

Given period is chosen as convenient (per minute; per hour; per day). Usually a variable, e.g. # customers arriving at a bank can be different between different 1-hour periods. Often seen to follow the Poisson Process, and is modeled thus.

P(r) = (re-) / r!
r: number of arrivals in the given time period. r = 0, 1, 2, .

: mean arrival rate (mean number of arrivals in a given time period).


P(r): the probability that exactly n arrivals will occur in the given period.

But, in general, arrival rate can follow any distribution.


Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Arrival Characteristics
INTER-ARRIVAL TIME (t)
The time elapsed between two consecutive arrivals into the system.

OR

Usually a variable, e.g. the time between two customers arriving at a bank can be different If the arrival behaviour follows a Poisson Process, then t follows the exponential distribution, and is modeled thus.

f(t) = e-t

: mean arrival rate = 1 / (mean of t).


f(t): the probability density function of t

ARRIVAL SIZE
Can be single arrivals, or batch arrivals
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Service Characteristics
SERVICE RATE
No. of customers that a server can handle in a given time period.

OR

For a given server, this can vary from one period to the next period. Hence, this is a random variable and has a distribution. As with the arrival process, the service process is often found to be Poisson. Multiple server system: distribution, its mean, variance can vary from server to server. i = 0, 1, 2, . m Pi(X) = (iXe-) / X!
X: number of customers served in the given time period. X = 0, 1, 2, . m: number of servers (channels).

i: mean service rate (the mean of X) of server i.


Pi(X): the probability that exactly X customers will be served by server i in the given period.
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Service Characteristics
SERVICE TIME

OR

Time taken by a server to handle a customer. For a given server, this can vary from customer to customer. Distribution can vary across servers. Multiple server system: distribution, its mean, variance can vary from server to server.

fi(t) = ie-it
i: mean service rate at server i = (mean service time or server i)-1.
fi(t): the probability density function of t

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Population

OR

Type of the population influences the arrival characteristics of the system.

FINITE POPULATION
Limited pool of customers. When customer leaves population, probability of next occurrence (or arrival rate) decreases, and vice versa. E.g. a small set of machines that need maintenance & service.

INFINITE POPULATION
Unlimited pool of customers. Customers leaving and rejoining the population does not influence arrival rate. E.g. the customer pool of a large bank, or a big retail store.
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Capacities
CAPACITIES

OR

a part of the system may be able to hold only up to a certain number of customers. particularly relevant in multi-phase systems, with finite capacities. e.g. the number of cars that can stand in line for drying after a carwash may be 10, hence car-wash has to become idle when the drying waiting line reaches 10.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Q discipline
A set of rules that determine the order of service offered to the customers in a given waiting line FIFO: First In First Out (FCFS: First Come First Served) LIFO: Last In First Out SIRO: Service in Random Order SPT: Shortest Processing Time First PR: Service according to priority Find real-life examples of each of these For all problems in this course, we assume the FIFO discipline
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

OR

QUEUEING SYSTEMS INPUTS


Inputs > Q behavior

OR

Do customers always wait in the line patiently? No, they also exhibit these behaviors Balking A customer does not enter the Q when he perceives its too long, and/or he does not have enough time to wait and avail the service. Reneging A customer waits in the Q for some time, but leaves when he sees it moving too slowly, and he does not have enough time. Jockeying A customer moves from his current line to another, e.g. if he feels he will receive service earlier by changing lines.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING SYSTEMS INPUTS


Inputs > Exit Fates
CUSTOMER REJOINS CALLING POPULATION
A serviced machine that may need the same service again.

OR

CUSTOMER DOES NOT REJOIN CALLING POPULATION


A cardiac patient passes away in a hospital. The machine overhauled, modified or repaired in such a way that there is a low probability of needing the same service again, or it is discarded.

Significantly influences the arrival characteristics of the system, when calling population is finite, but not when it is infinite.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

OR

TOGETHER, THE INPUT CHARACTERISTICS HELP DEFINE THE STRUCTURE OF THE QUEUEING SYSTEM

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

KENDALL QUEUEING NOTATION

OR

The Kendall notation has been widely adopted in queueing theory to depict the structure of the system in a simple manner. A/B/c/N/K
A: distribution of inter-arrival time. if exponential/markov, M. if constant/deterministic, D. if Erlang or order k, Ek if phase-type, P H if hyper-exponential, H if arbitrary, or general, G If general independent, G I B: distribution of service-time. (same notation as for A)
Operations Research

c: no. of channels / parallel servers. N: System capacity K: Size of the calling population.
Prof. Ajith Kumar

TAPMI, Theme 3, B2008, Feb-Apr 2009

KENDALL QUEUEING NOTATION


Example 1: M / M / 1 / /

OR

Single-channel (server) system; Both inter-arrival and service times follow exponential distributions; System capacity is unlimited; population is infinite

Example 2: G / G / 2 / 10 / 5
2 server system; Both inter-arrival and service times follow arbitrary distributions; System capacity is 10 customers; population has only 5 members
e.g. a set of 5 machines in a factory, maintained by 2 technicians, who can actually handle 10 machines with the resources available.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

KENDALL QUEUEING NOTATION


Simplified Kendall Notation

OR

When system capacity and calling population are infinite, they can be excluded from the notation. e.g. M / M / 1 / / can be written as M / M / 1 Many queueing systems can be approximated as having infinite calling population and system capacity; hence the simplified notation is popular.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

OR

NEXT, THE OUTPUTS

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

QUEUEING SYSTEMS OUTPUTS


Outputs: These are our performance measures.

OR

These not the inputs influence customer satisfaction / delight.


Long-run time-average of number of customers in the system (L) and the queue (Lq). Long-run average time spent by a customer in the system (W) and the queue (Wq). Long-run proportion of time, the system contains more than k customers. Server utilization, or the proportion of time the server is busy (). Server idle-time, or the proportion of time the no customer is in the system (Po). Q stability vs. instability.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

OR

STANDARD QUEUEING MODELS

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

STANDARD QUEUEING MODELS


1

OR

M/M/1: single-channel; exponential inter-arrival times & service times; infinite capacity & population

Important Assumptions Queue discipline: arrivals are served FIFO. Queue behavior: no balking/reneging; every arrival waits & receives service. Arrivals independent of each other; avg number of arrivals is constant over time. Arrival rate follows a Poisson distribution; and come from an infinite population. Service times vary from one customer to next & are independent; but their average rate is known. Average service rate () is > average arrival rate ().
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

STANDARD QUEUEING MODELS


1

M/M/1: formulas for the performance measures

OR

= Mean # arrivals in given time period, (or mean arrival rate) = Mean # customers served in given time period, (or mean service rate)
1. Avg. # customers in the system over time (L): 2. Avg. # customers in the queue over time (Lq): 3. Avg. time a customer spends in the system (W): 4. Avg. time a customer spends in the queue (Wq): 5. Utilization factor (): 6. % idle time; probability that no one in the system (Po): 7. Probability that no. of customers greater than k (Pn > k):
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009

L = / ( ) Lq = 2 / ( ) W = 1 / ( ) Wq = / ( ) =/ P0 = 1 ( / ) Pn > k = ( / )k+1

Prof. Ajith Kumar

STANDARD QUEUEING MODELS


2

M/M/m: multi-channel; exponential inter-arrival times & service times; infinite capacity & population

OR

Important Assumptions Same as for a single-channel queueing system.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

STANDARD QUEUEING MODELS


2

OR

M/M/m: formulas for the performance measures


m: number of channels (servers)

1. Avg. # customers in the system over time (L): L = P0 [ (/)m / (m-1)! (m )2 ] + / 2. Avg. # customers in the queue over time (Lq): Lq = L / 3. Avg. time a customer spends in the system (W): W=L/ 4. Avg. time a customer spends in the queue (Wq): Wq = W 1 / = Lq / contd
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009 Prof. Ajith Kumar

STANDARD QUEUEING MODELS


2

OR

M/M/m: formulas for the performance measures


m: number of channels (servers)

contd 5. Utilization factor (): = / m

6. % idle time; probability that no one in the system (Po): 1


n = m-1

(1/n!) (/)n n=0

for m >

+ (1/m!) (/)m (m/(m )

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

STANDARD QUEUEING MODELS


3

OR

M/D/1: single-channel; exponential inter-arrival times, constant service times; infinite capacity & population.
1. Avg. # customers in the system over time (L): 2. Avg. # customers in the queue over time (Lq): L = Lq + ( / ) Lq = 2 / 2( )

3. Avg. time a customer spends in the system (W): W = Wq + 1 / ( ) 4. Avg. time a customer spends in the queue (Wq): Wq = / 2( )

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

STANDARD QUEUEING MODELS


4

M / M / 1 / / K: single-channel, exponential inter-arrival time and service times, finite population of size K

OR

1. Avg. # customers in the system over time (L): L = Lq + (1 P0) 2. Avg. # customers in the queue over time (Lq): Lq = N [( + ) / ] (1 P0) 3. Avg. time a customer spends in the system (W): W = Wq + 1 / 4. Avg. time a customer spends in the queue (Wq): Wq = Lq / (K L) 1 5. % idle time; probability that no one in the system (Po): 6. Probability of n units in the system (Pn):
Operations Research TAPMI, Theme 3, B2008, Feb-Apr 2009

n=K

0[K!/(K-n)!] (/)n n=
P0 [K!/(K-n)!] (/)n
Prof. Ajith Kumar

STANDARD QUEUEING MODELS


General Relationships under Steady State Conditions
Applicable for any queue, under steady state (except the finite population model) L=W Lq = Wq (or W = L / ) (or Wq = Lq / )

OR

Littles Flow Equations

Also, avg. time in system = avg. time in q + avg. time in service W = Wq + 1 / The formulas are applicable under steady-state conditions.

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

OR
The preceding formulas are applicable under steady-state conditions, and under the respective assumptions made. What should we do when one or more of the assumptions do not hold and/or the queue in a transient state? SIMULATION MODELS ARE DEVELOPED FOR THIS

Operations Research

TAPMI, Theme 3, B2008, Feb-Apr 2009

Prof. Ajith Kumar

You might also like