Process Analysis

IE5107 OPERATIONS ANALYSIS AND MANAGEMENT

What is an excellent
production line?
STARTING QUESTION
Overview
Terminology and Little’s Law

Modeling and analysis of production processes

Workstations with exponential arrivals

 Q1-6

Terminology
Lead time (LT): The time from release of a job at the beginning of the routing until it reaches an
inventory point at the end of the routing.

LT = total process time + total waiting time

Capacity: Maximum production rate of a workstation or production process per unit time.
Terminology
Throughput (TH): Production rate of a workstation or production process per unit time.

TH = min {Arrival Rate, Capacity} Q7-9

Throughput of current workstation depends on the throughput of the previous workstation and
capacity of current workstation.

Q15
0.2 units 0.1 units 0.1 units
per hour per hour per hour
Terminology
Utilization (u): Fraction of time a workstation is not idling.

Q10-15

Note that u is a value between zero and one.

Work-in-progress (WIP): Number of jobs that are in the production line (i.e., jobs that have arrived and are not
completed)

The computation of WIP is more challenging as it changes over time (see Q16-31 of e-learning). For reporting purposes,
average WIP is often used.
What do managers like?
Lead time (short lead time)

Throughput (high throughput)

Utilization (high coz huge investment so want to ensure steady operating equipment)

WIP
Steady State
When arrival rate equals or is greater than capacity of production line:

◦ There is a bottleneck station where inventory builds up.

◦ Production line never enters steady state.

◦ WIP slowly increases over time (become infinity WIP over infinity time)

There is no meaningful average WIP to report.

Steady State
When arrival rate is less than capacity of production line:

◦ Production line enters steady state after some time (at the start, not really steady rate will be start up phrase)

◦ Average WIP (for stationary deterministic problem) can be computed by:

◦ Determining duration of steady state cycle (step 1)

◦ Observing proportion of time WIP is at a given level (step 2)

Q16-31

◦ Computing average WIP (step 3)

Observation: Computation of average WIP is more challenging than the computation of TH and CT.
Steady state

Next, we illustrate an easier

way to determine average WIP.

Observation: Computation of average WIP is more challenging than the computation of TH and LT.
Work-in-progress
Computation of average WIP is more challenging than the computation of TH and LT.

Little’s Law average WIP = TH LT Q32,33

Little’s Law provides an easy way to compute average WIP.

It also provide insights on production processes:

◦ Requiring high TH leads to high WIP (point 1) Next, we look at

how LT can be reduced.
◦ WIP can be lowered by reducing LT (point 2)
Observations
A shorter production lead time is preferred since it implies faster response to customer order and lower WIP.

LT = total process time + total waiting time What affects waiting time?
lower utilization under same demand potential for improvement
Lead time can be shorten by reducing process time (e.g., purchase faster machine).

A shorter process time implies higher capacity, which implies a lower utilization. -> hence trackle the waiting time instead
What do managers like (revisited)?
Lead time
WIP can be lowered by reducing LT => WT

Throughput

Utilization Requiring high TH leads to high WIP

WIP
Modeling production process
Previously, we considered the production process under stationary deterministic inter-arrival and
process times.

Next, we relax these assumptions to gain insights on waiting time for realistic problems.

Include data that manager likes: lead time/ average WIP/throughput/ utilization
Here, we consider three cases:
What data should we be collecting if we
◦ Single workstation want to analyze our process?

◦ Multiple servers What data should we be collecting if we

want to perform “what-if analysis”?
◦ Sequential workstations
Single workstation
Time
Job
IAT Process Arrival Start End
A 2 5
B 4 3
C 1 2
D 2 3
E 7 2

Arrival time = Previous arrival time + Inter-arrival time

Easy to Start time = max{Arrival time, Previous end time}
implement Q1-9
in MS Excel
End time = Start time + Process time
 Q10-11

Capacity
Time
Job
IAT Process Arrival Start End
A 2 5 2 2 7
B 4 3 6 7 10
C 1 2 7 10 12
D 2 3 9 12 15
E 7 2 16 16 18

Average process time min

Capacity min-1 hour-1 1
Capacity =
Process Time
 Q12-13

Throughput
Time
Job
IAT Process Arrival Start End
A 2 5 2 2 7
B 4 3 6 7 10
C 1 2 7 10 12
D 2 3 9 12 15
E 7 2 16 16 18

Average IAT min

Arrival rate min-1 hour-1
Throughput = min {18.75, 20} = 18.75 hour-1
TH = min \{Arrival Rate, Capacity \}
 Q15

Utilization
Time
Job
IAT Process Arrival Start End
A 2 5 2 2 7
B 4 3 6 7 10
C 1 2 7 10 12
D 2 3 9 12 15
E 7 2 16 16 18

Utilization
TH
u=
Capacity
 Easy to
implement in
Cycle time and wait time MS Excel

Time
Job
IAT Process Arrival Start End LT (End-arrival)
LT WTWT
(Start-arrival)

A 2 5 2 2 7 5 0
B 4 3 6 7 10 4 1
C 1 2 7 10 12 5 3
D 2 3 9 12 15 6 3
E 7 2 16 16 18 2 0

Average LT min 1.4 Proportion of WT/LT

×100 %=31.8 % -> 31.8% waiting time
4.4
Average WT min

Q16-22 LT = process time + WT

 Q23

Work-in-progress
Time
Job
IAT Process Arrival Start End LT WT
A 2 5 2 2 7 5 0
B 4 3 6 7 10 4 1
C 1 2 7 10 12 5 3
D 2 3 9 12 15 6 3
E 7 2 16 16 18 2 0

Average WIP Little’s Law

Average WIPwait
Single workstation analysis
We have illustrated how a single workstation can be modelled based on inter-arrival time and process time.

In addition, we illustrate how the following characteristics can be estimated:

◦ Capacity
◦ Throughput
◦ Utilization
◦ Average lead time
◦ Average wait time
◦ Average WIP
◦ Average number of items waiting

However, a production process may contain more than one workstation.

◦ Multiple servers
◦ Sequential workstations
 Suppose there
Multiple servers are 3 servers

Time
Job
IAT Process Arrival Start End
A 5 18 5 5 23
B 3 16 8 8 24
Q24-31
C 7 42 15 15 57
D 6 6 21 23 29
E 4 12 25 25 37

Arrival time = Previous arrival time + Inter-arrival time

Can we just look at the
Easy to Start time = max{Arrival time, large3{all previous end time}} previous three end times?
implement Cannot, in case if the processing time of previous 4 th
in MS Excel End time = Start time + Process time end time is too short so end time is lower
(n)th largest of previous (n) job So the large (nth) is not correct (maybe exam point->
test concept)
n depends on no of machine <- wrong concept
Sequential workstations
First Workstation Second Workstation
Job
I p a s e p a s e
A 2 5 2 2 7 1 7 7 8
B 4 3 6 7 10 7 10 10 17
C 1 2 7 10 12 2 12 17 19

Arrival time downstream = End time upstream

Start time = max{Arrival time, Previous end time}

Easy to
implement Q32-40
in MS Excel End time = Start time + Process time
Random number generation
We have illustrated how a production line can be modelled based on inter-arrival time and
process time.

We could simulate the production line by generating random inter-arrival time and process time
based on historical data.

Random numbers uniformly distributed between 0 and 1 can be easily generated in MS Excel.

◦ Data -> Data Analysis -> Random Number Generation

◦ RAND function
Random number generation
In practice, we may be interested in other distributions.
This can be achieved using the quantile function (inverse CDF).
−1
𝐹 𝑋 (𝑦)
CDF
3 Convert outcome to [0,1]

Quantile function
2 Convert [0,1] to outcome

𝑦
0.2 0.7 1
Quantile function (exponential distribution)
CDF of exponential distribution

−𝜆𝑥
Quantile function of exponential distribution

How is the quantile

function obtained?
𝑦=1−𝑒
Q1
Generating exponential random numbers
Suppose 0.095 is obtained from U(0,1) then the corresponding random number for exponential
distribution with parameter is:

Suppose jobs arrive at

a rate of 5 per hour.

In the e-learning exercise, we generate 25 exponentially distributed random numbers

Generating exponential random numbers

−1 − ln ( 1 − 𝑥 )
𝐹 ( 𝑥 )=
𝜆

How do we know it’s exponentially distributed?

 20% of values should be less than 0.0446

0.2 0.4 0.6 0.8
0.0446 0.1022 0.1833 0.3219
Generating exponential random numbers

−1 − ln ( 1 − 𝑥 )
𝐹 ( 𝑥 )=
𝜆

Sort in ascending order (show exponential random numbers and fit table result)

0.2 0.4 0.6 0.8

0.0446 0.1022 0.1833 0.3219
 Check distribution
 If exponentially distributed with , then 20% should be in each of the following bins:
• less than 0.0446
• [0.0446, 0.1022) Indeed, each bin had 5 out of
• [0.1022, 0.1833) the 25 “random numbers”
Q2-4
• [0.1833, 0.3219)
• at least 0.3219

 Pearson's chi-squared test (goodness-of-fit test) provides an objective way for assessing fit:

Q5-7

where is the number of bins used. As a guide, should be at least 5.

Chi-square table

d.f.= # bins – # est. parameters used in distribution – 1 Why need to “-1”

give some buffer/ “variation”
Degree of freedom = 5-2-1 to the dataset
Example: Check distribution
Suppose I have 25 numbers and I want to check if they are normally distributed

9.02 10.34 15.78 15.88 16.2 Mean: 21.54 SD: 5.83

16.99 17.83 18.12 19.35 20.62 20th percentile: 16.64
21.36 21.82 22.31 22.58 22.71
40th percentile: 20.07
23.34 23.81 24.59 24.72 25.01
60th percentile: 23.02
25.22 27.33 28.12 31.65 33.88
80th percentile: 26.45
Mean and variance of exponential distribution

Q7-8

1 1 How is the formula for

𝐸 [ 𝑋 ] = =0.2 𝑉𝑎𝑟 ( 𝑋 )= 2
=0.04 E[X] and Var(X) obtained?
5 5
Mean of exponential distribution
∞ ∞
𝐸 [ 𝑋 ]=∫ 𝑥⋅𝜆𝑒 𝑑𝑥=𝜆∫ 𝑥⋅𝑒 𝑑𝑥
− 𝜆𝑥 −𝜆𝑥

0 0
Variance of exponential distribution
∞ ∞
𝐸𝑉𝑎𝑟[ 𝑋( 𝑋 )=]=𝐸 [𝑋∫]−(𝑥𝐸[ 𝑋⋅𝜆𝑒 ∫
2 22 −𝜆𝑥 2 − 𝜆𝑥
2
]) 𝑑𝑥=𝜆 𝑥 ⋅𝑒 𝑑𝑥
0 0
M/M/1 queue
Next, we consider a workstation where:
◦ Inter-arrival time is exponentially distributed with parameter Assumption
◦ Process time is exponentially distributed with parameter

Suppose an average of 5 jobs arrive each hour and it takes an average of 6 minutes to process
each job:

Q9-10
M/M/1 queue
For a workstation with exponential inter-arrival and process time, the characteristics of the
process can be calculated as follows:
Worked
examples
provided in
Q11-16

Formulas are
verified through
simulation in
Q17 and Q18

Q11-18
M/G/1 queue
In practice, the processing time of the workstation may not follow the exponential distribution. Next, we consider a more
general problem:
◦ Inter-arrival time is exponentially distributed with parameter Assumption
◦ Process time follows a general distribution with mean and variance

Suppose an average of 5 jobs arrive each hour and it takes an average of 6 minutes to process each job. Furthermore, the
standard deviation of processing time is 3.464 minutes.

For example, process

time is uniformly
distributed between
0 and 12 min.
M/G/1 queue
For a workstation with exponential inter-arrival time, the characteristic of the process can be
calculated as follows:
Worked
examples
provided in
Q19-26

Formulas are
verified through
simulation in
Q27 and Q28

Q19-28
Waiting time analysis
Recall that reducing waiting time is desirable (i.e., reduce LT and WIP).

The formulas for average waiting time in M/M/1 and M/G/1 queues give us useful insights on
what affects waiting time.

M/M/1 M/G/1
Waiting time analysis
M/M/1 M/G/1

Observation 1:

◦ Waiting time in M/G/1 is merely waiting time in M/M/1 multiplied by a factor .

◦ In particular, waiting time increases with .

What is ?
 Q7 and Q8: The mean and
variance of an exponential
distribution with parameter λ

Coefficient of variation are λ-1 and λ-2, respectively.

Standard deviation is a measure of the spread in observations.

What is the CV of process time
However, the value of depends on the units used. For example: if it follows an exponential
distribution?

Coefficient of variation CV is a standardized measure of spread:

3.464 minutes
𝐶𝑉 = =0.58
6 minutes

CV provide a clearer description of variability.

Waiting time analysis
M/M/1 M/G/1

Observation 1:

◦ Waiting time in M/G/1 is merely waiting time in M/M/1 multiplied by a factor .

◦ Waiting time increases with coefficient of variation of workstation process time.

Waiting time analysis
M/M/1 M/G/1

Observation 2:

◦ Waiting time increases with . What is ?

◦ Waiting time increases with workstation utilization.

Waiting time analysis
M/M/1 M/G/1

Observation 3:

◦ Waiting time decreases with . What is ?

◦ If is large, then utilization is small.

◦ Hence, low utilization results in shorter waiting time.

◦ Waiting time increases with workstation utilization.

Waiting time analysis
Waiting time depends on the following two factors:

◦ Coefficient of variation of workstation

◦ Utilization of workstation

Wait
But managers like high utilization.
U
Critically, the relationship between the two is non-linear.
Reducing process time variability is
In general, 100% utilization is
vital for achieving high utilization
impossible to achieve in practice.
with reasonable waiting time.
 Any other ways
besides improving
Quick recap the workstation?
Sequential workstation
Jobs undergo process A followed by process B.
Charlie and Daniel have the same average process time for both processes.
Charlie’s process time is more variable than that of Daniel’s.

Min Max Average CV

Charlie 0 50 25 0.577
Daniel 20 30 25 0.115

Should we assign Charlie to process A or process B? Daniel will take the unassigned process.
Sequential workstation
Jobs undergo process A followed by process B.
Charlie works faster than Daniel on average.
Process time of Charlie and Daniel have the same CV.

Min Max Average CV

Charlie 0 20 10 0.577
Daniel 0 40 20 0.577

Should we assign Charlie to process A or process B? Daniel will take the unassigned process.
System design
In the previous slide, we have seen that “better workstations” should be positioned upstream
when possible.

What are better workstations?

◦ Faster
◦ Lower variability

Question: If a MRT station has 5 escalators leading to the train platform. How many the
escalators should be in the direction
◦ Towards train platform?
◦ Away from train platform Here, the variability is in the
arrivals, not the processing time.
First workstation
In some cases, the first workstation will process a job whenever it’s free rather than wait for a
job to arrive (i.e., no arrival rate).

How do we model such processes?

Workstation Workstation
1 2