Balancing Reduces Asymptotic Variance of Outputs
Download
Report
Transcript Balancing Reduces Asymptotic Variance of Outputs
A bit on Queueing Theory:
M/M/1, M/G/1, GI/G/1
Yoni Nazarathy*
EURANDOM, Eindhoven University of Technology,
The Netherlands.
(As of Dec 1: Swinburne University of Technology, Melbourne)
Swinburne University Seminar, Melbourne,
July 29, 2010.
*Supported by NWO-VIDI Grant 639.072.072 of Erjen Lefeber
Outline
•
•
•
•
•
•
The term: queueing theory
The single server queue
M/M/1, M/G/1, GI/G/1
Mean waiting time formulas
Derivation of the M/M/1 result
A glimpse at my queueing research
The Term: Queueing Theory
Queues
• Customers:
– Communication packets
– Production lots
– Customers at the ticket box
• Servers:
– Routers
– Production machines
– Tellers
• Queueing theory:
–
–
–
–
Quantifies waiting/congestion phenomena
Abstract models of reality
Mostly stochastic
Outputs:
• Performance evaluation (formulas, numbers, graphs)
• Design and control (decision: what to do)
Queueing Research
• 1909: Erlang – telephone lines
• Dedicated journal: Queueing Systems
• Other key journals:
–
–
–
–
–
•
•
•
•
Stochastic Models
Applied Probability Journals (JAP/Advances)
Annals of Applied Probability
OR, ANOR, ORL, EJOR…
About 5 other applied probability journals
Books: Around 200 Teaching/Research
Active researchers: ~500
Researchers that “speak the language”: ~2000
Related terms: “Applied Probability”, “Stochastic Modeling”
Queueing Theory Applied in Practice
• Here and there…
–
–
–
–
Practice motivates many new queueing problems
BUT: Queueing results not so often applied
Accurate data sometimes hard to obtain
Models are often too simple for very complex realities
• Simulation can do much more…
• …but say much less
• Insight gained from queueing theory is important
The Single Server Queue
The Single Server Queue
A Single Server Queue:
Server
Buffer
Number in
System:
Q(t )
0
1
2
3
4
5
6
…
Number in system at time t
Q(t )
t
The Single Server Queue
A Single Server Queue:
Server
Buffer
Number in
System:
Q(t )
0
1
2
3
4
5
6
…
Number in system at time t
{Tn , n 1} Arrivals times
{tn Tn Tn1, n 1} Inter-Arrivals times T0 0
{sn , n 1} Service requirements
The sequence
(tn , sn ), n 1 Determines evolution of Q(t)
Wn
The waiting time of customer n
Wn1 Max Wn tn1 sn ,0
(tn , sn ), n 1
Wn
Q(t )
Performance Measures
Assume the sequence is stochastic and stationary
E[tn ]
1
E[sn ]
Stable when
1
Load
1
Some important performance measures:
P(Wn x)
Little’s result:
W lim E[Wn ]
n
L lim E Q(t )
t
L W
We can quantify L (or W) under some further assumptions on
(tn , sn ), n 1
M/M/1, M/G/1, GI/G/1
Notation for Queues
• A/B/N/K
– A is the arrival process
– B is the service times
– N Is the number of servers
– K is the buffer capacity (default is infinity)
M/M/1, M/G/1, GI/G/1
(tn , sn ), n 1 :
Assumptions on
• M Poisson or exponential or memory-less
P X x e
• G General
• GI Renewal process arrivals
x
0
t
dt
Results for Mean Waiting Time
Mean Waiting Time
WM / M /1
1
1
2
1
c
s
WM / G /1 1
1 2
2
2 2
c
cs
1
a
WGI / G /1
1
2 2
ca
2
Var (tn )
E tn
2
, cs
2
Var (sn )
E sn
2
Derivation of the M/M/1 Result
A Markov Jump Process
Due to M/M (Exponential), at time t, Q(t) describes the state of the process
0
2
1
dP Q(t ) j
( ) P Q(t ) j P Q(t ) j 1 P Q(t ) j 1 , j 0,1, 2,...
dt
j lim P Q(t ) j
t
j
, j 0 Stationary distribution
L j j 1 0 Utilization
j 0
The Stationary Distribution
0
2
1
j j 1,
j 0
j
j 0,1, 2,...
1
Solution:
j (1 ) j
Performance measures:
L
1
1 0
lim P(Q(t ) j ) j
t
My Research
During PhD
• Control and stability of Queueing Networks
• Queueing Output Processes
2
2
2
c
c
1
s
K
Var DGI / G /1/ K , =1 (t ) a
lim
x
t
1 c 2 c 2 o (1) K
s
K
3 a
During Post-doc
• BRAVO Effect
(Seminar tomorrow at Melbourne University)
• Sojourn Time Tail Asymptotics
• Methods of Control Theory Applied to Queues
• Stability of Queueing Networks
• Asymptotic scaling of stochastic systems
• Optical Packet Switching Applications
In future…
• Research area: Model selection and statistics
of queueing networks (from data)
• Engineering applications
• More on previous subjects
• Power supply networks
Thanks for Listening and
See you Dec 1, 2010