Announcements

PS #3 out this afternoon Due: October 19 (graded by 10/23) Office hours Odoni: Mon. 2:30-4:30
- Wed. 2:30-4:30 on Oct. 18 (No office hrs 10/16) _ Or send me a message

Queueing Systems: Lecture 1

Amedeo R. Odoni
October 4, 2006

Quiz #1: October 25, open book, in class Old quiz problems and solutions:

posted on 10/19

Topics in Queueing Theory

Lecture Outline
Introduction to queueing systems Conceptual representation of queueing

Introduction to Queues Littles Law Markovian Birth-and-Death Queues The M/M/1 and Other Markovian Variations The M/G/1 Queue and Extensions Priority Queues Some Useful Bounds Congestion Pricing Queueing Networks; State Representations Dynamic Behavior of Queues

systems Codes for queueing models Terminology and notation Littles Law and basic relationships Birth-and-death models The M/M/1 queueing system
Reference: Chapter 4, pp. 182-203

Queues
Queueing theory is the branch of operations research

A Generic Queueing System

Servers Arrival point at the system Queue

concerned with waiting lines (delays/congestion) A queueing system consists of a user source, a queue and a service facility with one or more identical parallel servers A queueing network is a set of interconnected queueing systems Fundamental parameters of a queueing system: - Demand rate - Capacity (service rate) - Demand inter-arrival times - Service times - Queue capacity and discipline (finite vs. infinite; FIFO/FCFS, SIRO, LIFO, priorities)
- Myriad details (feedback effects, balking,
jockeying, etc.)

C C C C C C C

Departure point from the system

Source of users/ customers

C C CC CC

Arrivals process Size of user source Queue discipline and Queue capacity Service process Number of servers

Queueing network consisting of five queueing systems

Applications of Queueing Theory

Some familiar queues: _ Airport check-in; aircraft in a holding pattern _ Automated Teller Machines (ATMs) _ Fast food restaurants _ Phone centers lines _ Urban intersection _ Toll booths _ Spatially distributed urban systems and services Level-of-service (LOS) standards Economic analyses involving trade-offs among

Queueing system 2 Queueing system 1 Point where users make a choice Queueing system 4

Queueing system 3 Point where users merge Queueing system 5 Out

In

operating costs, capital investments and LOS Congestion pricing

The Airside as a Queueing Network

Queueing Models Can Be Essential in

Analysis of Capital Investments

Cost Total cost

Optimal cost

Cost of building the capacity

Cost of losses due to waiting

Optimal capacity

Airport Capacity

Strengths and Weaknesses of

Queueing Theory

Queueing models necessarily involve approximations

A Code for Queueing Models:

A/B/m

Distribution of service time Queueing System Number of servers Customers Queue

and simplification of reality Results give a sense of order of magnitude, changes relative to a baseline, promising directions in which to move Closed-form results essentially limited to steady state conditions and derived primarily (but not solely) for birth-and-death systems and phase systems Some useful bounds for more general systems at steady state Numerical solutions increasingly viable for dynamic systems

//
Distribution of interarrival time

CCCCCC

C C C C

S S S S

Service facility

Huge number of important applications

Some standard code letters for A and B: _ M: Negative exponential (M stands for memoryless) _ D: Deterministic _ Ek:kth-order Erlang distribution _ G: General distribution

Terminology and Notation

Number in system: number of customers in

Terminology and Notation (2)

Transient state: state of system at t is

queueing system Number in queue or Queue length: number of customers waiting for service Total time in system and waiting time N(t) = number of customers in queueing system at time t Pn(t) = probability that N(t) is equal to n at time t n: mean arrival rate of new customers when N(t) = n n: mean (total) service rate when N(t) = n

influenced by the state of the system at t = 0

Steady state: state of the system is independent

of initial state of the system m: number of servers (parallel service channels) If n and the service rate per busy server are constants and , respectively, then n=, n= min (n, m); in that case:
_ Expected inter-arrival time = 1/
_ Expected service time = 1/

Some Expected Values of Interest at Steady State

Given:
_ _
Number of users

Littles Law
A(t): cumulative arrivals to the system C(t): cumulative service completions in the system

= arrival rate
= service rate per service channel

A(t)

Unknowns:

L = expected number of users in queueing system _ Lq = expected number of users in queue _ W = expected time in queueing system per user (W = E(w)) _ Wq = expected waiting time in queue per user (Wq = E(wq)) 4 unknowns We need 4 equations
_

N(t) C(t)

Time

LT = 0

N (t ) dt T

A(T ) 0 T A(T )

N (t ) dt

= T WT

Relationships among L, Lq, W, Wq

Four unknowns: L, W, Lq, Wq Need 4 equations. We have the following 3 equations:
_ _ _

Birth-and-Death Queueing Systems

1. m parallel, identical servers. 2. Infinite queue capacity (for now). 3. Whenever n users are in system (in

L = W (Littles law) Lq = Wq W = Wq +
1

If we can find any one of the four expected values, we

can determine the three others

The determination of L (or other) may be hard or easy

depending on the type of queueing system at hand

L = nPn (Pn : probability that n customers are in the system)

n=0

queue plus in service) arrivals are Poisson at rate of n per unit of time. 4. Whenever n users are in system, service completions are Poisson at rate of n per unit of time. 5. FCFS discipline (for now).

The Fundamental Relationship

Time: t
nt
1-(n+ n)t

The differential equations that determine the state probabilities

Pn (t + t) = Pn +1 (t) n +1 t + Pn 1 (t) n 1 t + Pn (t) [1 ( n + n ) t]

Time: t+t n+1 users Pn(t) = Prob [n users in system at time t] n users

After a simple manipulation:

dPn (t) = (n + n ) Pn (t) + n 1 Pn 1 (t) + n +1 Pn +1 (t) dt

(1) applies when n = 1, 2, 3,.; when n = 0, we have:
dP0 (t) = 0 P0 (t) + 1 P (t) 1 dt

(1)

n users
nt

(2)

n-1 users
Pn (t + t) = Pn +1 (t) n +1 t + Pn 1 (t) n 1 t + Pn (t) [1 ( n + n ) t]

The system of equations (1) and (2) is known as the Chapman-Kolmogorov equations for a birth-and-death system

The state balance equations

We now consider the situation in which the queueing

Birth-and-Death System: State Transition Diagram

0
0 1

system has reached steady state, i.e., t is large enough to have Pn (t) = Pn , independent of t, or dPn (t) = 0 dt
Then, (1) and (2) provide the state balance equations:

1
2

m-1
m

m
m+1

m+1

m+1

m+2

0 P0 = 1 P1 (n + n ) Pn = n 1 Pn 1 + n +1 Pn +1
from the state transition diagram

n=0

(3)

n = 1, 2, 3,.. (4)

The state balance equations can also be written directly

We are interested in the characteristics of the system under equilibrium conditions (steady state), i.e., when the state probabilities Pn(t) are independent of t for large values of t Can write system balance equations and obtain closed form expressions for Pn, L, W, Lq, Wq

Solving..
Solving (3) and (4), we have:

M/M/1: Observing State Transition Diagram from Two Points

From point 1:
( + )Pn = Pn1 + Pn+1

P = 0 P0 ; P2 = 1 P = 1 0 P0 1 1 1 2 2 1
and, in general, ..... 1 0 P0 = K n P0 Pn = n 1 n 2 n n 1 ..... 2 1 But, we also have: Giving,

etc.

P0 = P ( + )P = P0 + P2 1 1

0 1

n-1 n

n+1

1=

n =0

Pn = P0 (1 + K n )
n =1

From point 2:
P0 = P 1

P1 = P2

1 2

Pn = Pn+1

n-1 n

P0 = 1+

1
n =1

Kn

Condition for steady state: Kn < n=1

n+1

M/M/1: Derivation of P0 and Pn

P0 , P2 = P0 , L, Pn = P0

M/M/1: Derivation of L, W, Wq, and Lq

L = nPn = n n (1 ) = (1 ) n n = (1 ) n n1
n=0 n=0 n=0 n=1

Step 1:

P= 1

Step 2:

P
n=0

= 1, P0 = 1 P0 = n=0

d n d 1 = (1 ) = (1 ) d n=0 d 1
n

1
n=0

Step 3:

, then
P0 =

=
n=0

n=0

1 1 = (Q < 1) 1 1

1 = = (1 ) (1 ) 2 = (1 ) = 1 L 1 1 W= = =

1 Wq = W =

( )

Step 4:

n=0

= 1
n

and Pn = n (1 )

Lq = Wq =

2 = ( ) ( )

High Sensitivity of Delay at High

Levels of Utilization

Expected delay

M/M/1: An alternative, direct derivation of L and W

For an M/M/1 system, with FCFS discipline:

W =
n=0

(n +1)

Pn = E[

N +1

(2)

]=

E[N ] +1

L +1

(1)

But from Littles theorem we also have:

L = W
Demand Capacity

It follows from (1) and (2) that, as before:

L=

W =

=1

Does the queueing discipline matter?

Additional important M/M/1 results

The pdf for the total time in the system, w, can be computed for a M/M/1 system (and FCFS):

M/M/1: E[B], the expected length of a

busy period

B = busy period I = idle period

f w (w) = (1 ) e (1 ) w = ( )e ( )w for w 0
Thus, as already shown, W = 1/( -) = 1/[ (1-)] The standard deviation of N, w, Nq, wq are all proportional to 1/(1-), just like their expected values (L, W, Lq, Wq, respectively) The expected length of the busy period, E[B], is equal to 1/( -)
I B I B I

P0 =
But,

E[length Idle period ] E[length Busy period ] + E[length Idle period ]

P0 = 1 E[length Idle period ] = 1
1

1 = 1

Therefore,

E[B] = E[length Busy period ] =

(1 )