Tip:
Highlight text to annotate it
X
In the previous session, we saw that waiting time can occur even if the
resource utilization is below 100%.. This was driven by the variability in the
process flow. Now unfortunately, the type of tools that
we've introduced so far in this course, can really not explain that waiting time.
We, so far naively believed that if the utilization is less than 100%,, there
wouldn't be any waiting. The reason for this is our previous
analysis was always built on averages. We have simply ignored the concept of
variability, so far. The purpose of this session is to help us
describe the variability in this process, and start putting together a set of tools
that help us predict the amount of waiting, even if the utilization is less
than 100%.. Now, we first need to describe the
variability, we need some form of measuring it.
We'll stay with the example of the doctor's office, this is pretty
representative for most service operations.
I have some waiting area and I have a resource.
Notice that in the section now, we're going to look at the case where the flow
rate is constrained by demand. If we have more demand than capacity, our
implied utilization is more than a 100% and we don't need fancy variability models
to tell us that there's waiting. This was my first doctor's office in the
previous session. We further assume for the remainder of
this session that everybody who arrives to the practice will wait in line till they
are served and leave the patients after having seen the doctor.
Okay. Now, let's first describe variability in
demand. We notice that the demand process was
somewhat random. When we said random, what we meant is that
the customers were not lined up at Toyota Camrys at the end of an assembly line.
They came when they wanted. So, a formal way of capturing this is we
define the arrival time as a time when patients arrive to the practice.
And then we define the inter-arrival time, as the time between two subsequent
arrivals. When we say that the arrivals are random,
what we really mean is that these inter-arrival times are drawn from some
underlined statistical distribution. We will denote with a, the average
inter-arrival time. Like any distribution, the inter-arrival
time distribution not just has an average but also a standard deviation.
We denote by the coefficient of variation of the inter-arrival time, CVa, as the
ratio between the standard deviation and the mean.
We call the idea of the coefficient of variation from the module of customer
choice and variety. The idea here is that the standard
deviation by itself is not a good measure of variability.
Is a standard deviation of ten minutes a lot or a little?
Well, it really depends on the underlying demand process.
You might have heard concepts such as exponential distribution or Poisson
distributions. For those cases, the CVa is equal to one, and the assumption is
really that at any given moment in time, the likelihood of a new patient arriving
in the next unit of time is constant. I'll talk about seasonal arrival times a
little later on in this module. For now, I'll just assume that the
likelihood of a patient coming in is random, but is reasonably constant over
time. Variability however, is not limited to the
demand process. We also have variability in what we do
internally. In other words, we have variability in our
processing time. Let me use P to define the average
processing time, Keeping in mind that the processing time
will vary from one customer to the other. Just like we used the coefficient of
variation for the inter-arrival times, it also defines a coefficient of variation
for the processing time, as a standard deviation of the processing times divided
by the average processing time. The coefficient of variation of the
processing time really measures how standardized the work is.
This varies a lot by application. In an assembly line forever, you might see
situations where this coefficient of variation of the processing time might be
close to zero. In other settings, if you think about an
operating room, you think about developing a piece of software, there might be a fair
bit of variability in the processing time. Hence, CVps can be one or bigger.
Again, this is something that you would have to measure as you go into the
analysis that we're about to do. You just collect a bunch of processing
times, put them in Excel, And just compute the standard deviation
and the mean. Equipped with all of these definitions, we are now ready to compute
the time in the queue. Notice that the time in the queue will
vary for each individual customer. So, the best we can do, is we can compute
the expected or the average time in the queue.
Now, the following formula will allow us to compute this time in the queue.
The time in the queue is given as a product of the average processing time P
for the average activity time, I use those interchangeably here,
Times the utilization divided by one minus the utilization times the coefficients of
variations added up and then divided by two.
Now, let me comment on this and go to each of these three factors step by step.
The first one is simply the average processing time.
In other words, we see that the time in the que grows linearly here with the
processing time P. Whenever in math you see a formula that
has the shape u divided by one minus u, you know that it gets ugly as you're
approaching one, then we just plug in some numbers. At an 80% utilization,
This ratio is 0.8 divided by 0.2 which is simply four.
At a 90% utilization, The same ratio is 0.9 divided by 0.1 which
is already nine. You notice that a simple ten percent
increase in utilization can more than double the waiting time.
You see this on the graph here on the right that the utilization grows very
steeply as we approach 100% utilization. This is practically very important.
The reason for that is that managers and service operations are incurring very big,
fixed costs. So, it is in their interest to squeeze
more and more customers through the process.
However, you'll notice that those last customers, which from a profitability
perspective are really interesting because their revenue goes right into the bottom
line after all the fixed costs are already paid for.
They look very profitable, but they create havoc to the system.
Again, as these last customers get over to the process, our waiting times get up to
the roof. Finally, look at the variability here.
I'm squaring the coefficient of variation for both the arrival process as well as
for the processing times. The more variability there is in the
process, the longer is the time in the queue.
If you look at the case where you have a coefficient of variation of a exponential
distribution, you notice that these two fellows here are going to be equal to one,
and this whole loss factor will degenerate to one.
Let's practice our new equation with a quick example.
Again, let's look at a doctor's office. Let's assume that patients come every 30
minutes, or is with a standard deviation of 30 minutes, and that consultations last
fifteen minutes with a standard deviation of fifteen minutes.
The first thing that I encourage you to do for these sets of problems, is simply
write down the waiting time formula. The time in the queue is P divided by u
divide by one minus u times the coefficients of variations added up after
they got squared individually. Now, which one of these ingredients here
in the formula is easy? Well, we see that the processing time is
simply fifteen minutes. What else? The coefficient of variation of
the inter-arrival time is 30 minutes off the standard deviation divided by 30
minute average. So, this means it's one, which gets
squared, but it still stays one, same for the processing times, the standard
deviation here is fifteen minutes, the average is fifteen minutes, so I have
another one squared. And I divide those two by two.
So, this old fellow here at the end is simply going to be equals to one.
Now, utilization is not immediately visible in this question, and so we have
to remind ourselves that utilization is the flow rate divided by the capacity.
Flow rate in this practice is one patient every 30 minutes unconstrained by demand.
I divide this by capacity, which is one over fifteen, then the utilization of
50%.. In other words then, if I plug this in,
this middle factor here is simply 50% divided by one minus 50%, which is equals
to one. So, the total wait time does, is fifteen
times one times one equals to fifteen minutes.
Now, I want to be a little hairsplitting here, because the question actually
doesn't ask for the waiting time, but asks what's the time when our friend, Newt, can
walk out of the practice again. That includes the time in the queue,
Which we said was fifteen minutes, but also the time in the practice when he is
in service. And so, this is the time in the queue the
plus the processing time, Altogether, 30 minutes.
And so, at ten:30,30, he can expect to be out of the office again.
We can predict the average time in the queue, based on the processing time, the
utilization, and the amount of variability, in the system.
We saw on the waiting time formula that as the utilization goes up and it approaches
100%, the utilization will drive the waiting time through the roof.
So, once the system becomes more congested than a 90, 95%, utilization, it becomes
very sensitive to every additional customer walking in.
Please only use this waiting time formula for situations where the utilization is
less than 100%, or we have more capacity than we have demand.
If demand exceeds capacity, first of all, remember, we speak of an implied
utilization, utilization by itself at the maximum be 100%..
But if demand exceeds capacity we really don't have a variability in dues waiting
time. We just have a problem of a doctor who is
taking six patients in and can only serve three.
To predict that this waiting room will fill up over time, doesn't require some
fancy math. It's just a simple matter of understanding
that we are serving three fewer customers than we have demand.
Also notice that once we have computed the time in the queue, we can compute the
total flow time, the time in the system, by simply adding the time in the queue
plus the processing time P. We can also compute inventory using
Little's Law. So, once you have computed the time in the
queue, every other measure that we talked about in this class can be computed.
It is why this waiting time formula is so important.