Capacity Management
INFO 564 Spring 2020
Long-Run Capacity Management
• Long-run capacity management decisions are based on business strategy
• What business will we be in?
• What does the future look like – technology, competitors, sustainability?
• Markets and customers?
• How will we compete - cost, differentiation?
• Tools and techniques
• R&D
• Long-term forecasting; technological forecasting
• Market research
• SWOT analysis
• Decision anlaysis
• Decisions about: aggregate capacity in chosen sectors; geographical
location; investments in technology and expertise
Capacity Management in the Medium Term
• In the medium term, strategy, technology, markets, and products tend to
be fixed
• Key uncertainty is demand and actions of competitors
• Capacity management strategies
• Lead strategy – have more capacity than needed in anticipation of demand
• Lag strategy – add capacity only after demand materializes
• Techniques: forecasting, decision trees, optimization
• Decisions about
• How much capacity to have (production capacity, workforce)
• Leasing, outsourcing
• Location, especially with services
Capacity Management in the Short Run
• In the short run almost everything is fixed: capacity, demand, technology,
competition.
• What is variable is the timing of demand.
• To match capacity with demand: overtime, inventory, extra shifts, hiring
and layoff, temporary workers
• Short-run capacity management is a critical function of operations
management.
• Services pose a special challenge
• Timing of demand can be highly variable; some variation can be anticipated
• Service cannot begin until people arrive; often depends on their input
• People cannot be made to wait too long
Short-Run Capacity Management in Services
• Techniques
• Optimization - tries to allocate capacity so that some measure of efficiency is
optimized – minimize labor costs
• Queueing Theory - looks at the interaction of service capacity and demand;
provides insight into customer waiting and server utilization.
Queueing Theory
• Demand – expressed as an average arrival rate with a probability
distribution
• Customers per hour, ships per day, patients per week, etc.
• Probability distribution often assumed to be Poisson
• Capacity – expressed in terms of
• An average service rate per server with a probability distribution
• Customers per hour, ships per day, patients per week, etc.
• # of servers
• Probability distribution of service rate often Poisson
A Basic Queueing System
Served Customers
Queueing System
Queue
C S
Customers CCCCCCC C S Service
C S facility
C S
Served Customers
Common Queuing System Configurations
CustomerArrives CustomerLeaves
...
Waiting Line Server
CustomerLeaves
Server 1
CustomerArrives CustomerLeaves
...
Waiting Line Server 2
CustomerLeaves
Server 3
CustomerLeaves
...
Waiting Line Server 1
CustomerArrives CustomerLeaves
...
Waiting Line Server 2
... CustomerLeaves
Waiting Line Server 3
Model Notation
• Queuing systems are described by 3 parameters: 1/2/3
• Parameter 1
M = Markovian interarrival times
D = Deterministic interarrival times
• Parameter 2
M = Markovian service times
G = General service times
D = Deterministic service times
• Parameter 3
A number Indicating the number of servers.
• Examples: M/M/3, M/G/4, M/G/2
Queueing Models
• Infinite Queue Models
• M/M/S
• First M: Demand rate is Poisson distributed
• Second M: Service rate is Poisson distributed
• S: # of servers
• No limit on the number of people who can wait in the line
• Finite Queue Models
• M/M/S finite queue
• Limit on the number of people who are allowed to wait in line
• Finite Population Models
• M/M/S finite population
• Limited demand. While all that demand is waiting to be satisfied, there can be no
more demand.
Queueing Models Give Us:
L = Expected number of customers in the system, including those being
served (the symbol L comes from Line Length).
Lq = Expected number of customers in the queue, which excludes customers
being served.
W = Expected waiting time in the system (including service time) for an
individual customer (the symbol W comes from Waiting time).
Wq = Expected waiting time in the queue (excludes service time) for an
individual customer.
These definitions assume that the queueing system is in a steady-state
condition.
11.11
Choosing a Measure of Performance
• Managers who oversee queueing systems are mainly concerned with
two measures of performance:
• How many customers typically are waiting in the queueing system?
• How long do these customers typically have to wait?
• When customers are internal to the organization, the first measure
tends to be more important.
• Having such customers wait causes lost productivity.
• Commercial service systems tend to place greater importance on the
second measure.
• Outside customers are typically more concerned with how long they have to
wait than with how many customers are there.
The M/M/s Model
• Assumptions:
• There are s servers.
• Arrivals follow a Poisson distribution and occur at an average
rate of l per time period.
• Each server provides service at an average rate of m per time
period, and actual service times follow an exponential
distribution.
• Arrivals wait in a single FIFO queue and are serviced by the first
available server.
• l< sm.
An M/M/s Example: Bitway Computers
• The customer support hotline for Bitway Computers is currently staffed by a single
technician.
• Calls arrive randomly at a rate of 5 per hour and follow a Poisson distribution.
• The technician services calls at an average rate of 7 per hour, but the actual time
required to handle a call follows an exponential distribution.
• Bitway’s president, Rod Taylor, has received numerous complaints from customers
about the length of time they must wait “on hold” for service when calling the
hotline.
• Rod wants to determine the average length of time customers currently wait
before the technician answers their calls.
• If the average waiting time is more than 5 minutes, he wants to determine how
many technicians would be required to reduce the average waiting time to 2
minutes or less.
• A technician is paid $18 an hour. Imputed customer waiting cost is $30 per hour
(in lost goodwill, etc.) On the basis of hourly total cost what is the right number of
technicians to have?
Relationship between L, W, Lq, and Wq
• Since 1/m is the expected service time
W = Wq + 1/m
• Little’s Law states that
L = lW
and
Lq = lWq
• Combining the above relationships leads to
L = Lq + l/m
Using Probabilities as Measures of
Performance
• In addition to knowing what happens on the average, we may also be interested
in worst-case scenarios.
• What will be the maximum number of customers in the system? (Exceeded no more than,
say, 5% of the time.)
• What will be the maximum waiting time of customers in the system? (Exceeded no more
than, say, 5% of the time.)
• Statistics that are helpful to answer these types of questions are available for
some queueing systems:
• Pn = Steady-state probability of having exactly n customers in the system.
• P(W ≤ t) = Probability the time spent in the system will be no more than t.
• P(Wq ≤ t) = Probability the wait time will be no more than t.
• Examples of common goals:
• No more than three customers 95% of the time: P0 + P1 + P2 + P3 ≥ 0.95
• No more than 5% of customers wait more than 2 hours: P(W ≤ 2 hours) ≥ 0.95
The M/G/1 Model
• The exponential distribution assumes a high amount of variability – the
standard deviation is equal to the mean. It is a “worst case” distribution.
• Not all service times are this variable and so the exponential is not the
appropriate distribution for them.
• Changing oil in a car, getting an eye exam, getting a hair cut, etc.
• A different model may be used in these cases – M/G/1 model
• M/G/1 Model Assumptions:
• Arrivals follow a Poisson distribution with mean l.
• Service times follow any distribution with mean m and standard deviation s.
• There is a single server.
An M/G/1 Example: Zippy Lube
• Cars arrive at the Zippy-Lube oil change center following a Poisson
distribution at an average rate of 3.5 cars per hour.
• The average service time per car is 15 minutes (or 0.25 hours) with a
standard deviation of 2 minutes (or 0.0333 hours).
• A new automated oil dispensing device exists
• The manufacturer's representative claims this device will reduce the
average service time by 3 minutes per car. (Currently, employees
manually open and pour individual cans of oil.)
• The owner wants to analyze the impact the new automated device
would have on his business.
Finite Queue M/M/S Models
• Finite Queue: the queue can only hold a limited number of
people.
• Entities cannot join the line after it has reached its limit. The
arrival rate becomes zero.
• Many queues are finite queues
• Physical space limitations
• Planes waiting to land; waiting to take off; ships waiting to
unload; phone systems that can keep a limited number on hold.
Example
Road Rambler sells specialty running shoes. Customers can phone in orders at any
time day or night, 7 days a week. One rep handles calls which arrive at a rate of 14
per hour Poisson distributed. It takes the rep an average of four minutes to process
each call, exponentially distributed. Suppose that Road Rambler’s phone system
can only keep four calls on hold at any time; after that callers receive a busy signal.
The average profit margin of each call is $55, and sales reps cost the company $12
per hour.
a. If callers who receive a busy signal take their business elsewhere, how much
money is the company losing per hour (on average) if it employed a single sales
rep?
b. What is the net effect on average hourly profits if the company employs two
sales reps instead of one?
c. How many sales reps should the company employ if it wants to maximize profit?
11.20
Finite Population M/M/S Model
• Limited population of customers
• Dependent relationship between queue length and arrival rate
• Assumptions
1. There are s servers with identical service time distributions.
2. The population of units seeking service is finite, of size N.
3. The arrival distribution of each customer in the population follows a Poisson
distribution, with an average rate of l.
4. Service times are exponentially distributed, with an average rate of m.
5. Both l and m are specified for the same time period.
6. Customers are served on a first-come, first-served basis.
Example
• DeColores Paint Company owns 10 trucks that it uses to deliver paint and deco-
rating supplies to builders. On average, each truck returns to the company’s
single loading dock at a rate of three times per 8-hour day. The times between
arrivals at the dock follow an exponential distribution. The loading dock can
service an average of 4 trucks per hour with actual service times following an
exponential distribution.
a. What is the probability that a truck must wait for service to begin?
b. On average, how many trucks wait for service to begin at any point in time?
c. On average, how long must a truck wait before service begins?
d. If the company builds and staffs another loading dock, how would your answers to parts a, b,
and c change?
e. The capitalized cost of adding a loading dock is $5.40 per hour. The hourly cost of having a
truck idle is $50. What is the optimal number of loading docks that will minimize the sum of
dock cost and idle truck cost
11.22
Some Insights About Designing Queueing
Systems
1.When designing a single-server queueing system, beware that giving
a relatively high utilization factor (workload) to the server provides
surprisingly poor performance for the system.
2.Decreasing the variability of service times (without any change in the
mean) improves the performance of a queueing system substantially.
3.Multiple-server queueing systems can perform satisfactorily with
somewhat higher utilization factors than can single-server queueing
systems. For example, pooling servers by combining separate single-
server queueing systems into one multiple-server queueing system
greatly improves the measures of performance.
11.23
Managing Process Flows
(Summary of Note on Managing Process Flows – Ramdas)
Capacity
• The number of items that can be processed in a given amount of
time.
• Bottleneck capacity
• Cycle time
• How much of this capacity is utilized is measured by capacity
utilization
Throughput Time
• How quickly an item (flow unit) gets through a process
• For a particular item it is the difference between the time it exits the
process and the time it entered the process.
• In real life processes, different items can take different amounts of
time to go through the system.
• Thus we can think of an average throughput time.
Throughput Rate
• The rate at which units exit the process
• Measured as units/time-unit
• Customers/hour
• Repairs per week
• Landings per day
• Etc.
• This rate can be variable, so in practice we can talk about an average
throughput rate.
• This is the same as the arrival rate in queueing theory.
Inventory
• Inventory denotes items in process or waiting to be processed.
• This number changes with time
• Units enter the process
• Units leave the queue to get processed
• Units get processed and leave the process
• So we can think in terms of average inventory
Equilibrium or Steady State
Arrival
Service Rate μ Exit Rate λ
Rate λ
For steady state the arrival rate has to be smaller than the service rate.
When this prevails, the rate at which units leave the service after being
served is the same as the arrival rate.
Some Insights from Little’s Law
• L = λW
• Average Number in System = Average Arrival Rate * Average Time in
System.
• Average Number in System maybe referred to as Average Inventory,
Average Arrival Rate as Throughput Rate and Average Time in System
as throughput time,
• Average Inventory = Average Throughput Rate * Average Throughput
Time
• I = RT
Examples
• Suppose customers arrive at a doctor’s office at an average rate of 10
per hour and on average customers spend 75 minutes (waiting and
getting treated).
• Here R = 10 customers per hour, T = 1.25 hours. So I = R*T = 10*1.25
or 12.5 customers on average.
• In other words we can expect on average 12-13 customers to be in
this office. Some of these customers are waiting, others are in the
process of being treated.
• The office has to make arrangements to accommodate this many
customers.
Insurance Example – Current Configuration
I = 29.9*10 = 299 claims
R = 29.9/day Under- Policy
Registration 29.9/day 29.9/day Rating 29.9/day R = 29.9/day
writing Writing
T = 10 Days
How to reduce the throughput time?
Alternate Configuration For standard claims throughput time is the sum
of: Registration (60/29.9) + StdUW (40/23.9) +
Rating (50/29.9) + PW (30/29.9) = 2 + 1.7 + 1.7
+ 1 = 6.4 days
Standard
Underwriting
I=40
Policy
R = 29.9/day Rating 29.9/day R = 29.9/day
Registration Writing
I=50 I=30
I=60
Complex
Underwriting
I=20 For complex claims throughput time is the sum
of: Registration (60/29.9) + CompUW (20/6) +
Overall, I = 200 claims. Rating (50/29.9) + PW (30/29.9) = 2 + 3.3 + 1.7
R = 29.9/day + 1 = 8 days
T = 200/29.9 = 6.7 days
Inventory and COGS
COGS = $12 Production Process COGS = $12
bn/yr I=$1b bn/yr
Cost of Goods Sold may be interpreted as Throughput Rate per year, R = $12
billion.
I = $1 billion
Throughput Time = I/R = 1/12 years.
This means that this process runs through its inventory 12 times each year. This
is referred to as inventory turns.
Recap
• Little’s Law applies to any queueing system in steady state.
• Does not depend on probability distributions of arrivals or service.
• Applies to single server or multiple server systems.
• Applies to single-phase as well as multiple-phase systems
• Gives us a quick way of estimating basic characteristics of processes
• Knowing 2 of the 3 numbers allows us to calculate the 3rd.
• Allows us to make quick assessments of inventory and throughput on
the basis of demand.