Network Traffic Projection Method
Overview
The network traffic projection method described
below is easy to use and remarkably accurate.
In a nutshell the method uses historical network traffic and headcount
data to project future traffic growth.
Those two sources of data are used to derive the traffic per user. The average growth or decline in traffic per
user is applied to future headcount projections and the result is a fairly
reasonable idea of future traffic levels.
The first year this method was employed at the large corporation I work
at it came within 3.8% of actual traffic levels, on a network moving more than
100,000 Terabytes a year (yes it is true that the number just mentioned could
be also be called 100 Petabytes, but for some reason many people are
uncomfortable with that term).
This approach altered how we think about traffic
loads at my company. Empirical evidence
from the past seemed to imply that headcount was not that big a factor in
projecting future network loads because use was growing exponentially
regardless of hiring or firing. This was
the case where I work from the early 90’s to 2000; network traffic was always
increasing regardless of employment levels.
Traffic growth started to change in 2001, which is when this method was
devised. Our traffic levels started to
level off and there were announcements of some pretty heavy layoffs
coming. At the same time my company had
reached computing saturation, after years of expanding the network in every
direction all the time, everyone who needed a networked computer finally had
one and most if not all the software they needed. This got me to thinking that things had
evolved enough that large changes in headcount might indeed affect future
traffic levels.
Using data collected from 1997 through 2001 this
method was applied and averaged 5% accuracy per year over that five-year
period, a period of substantial growth.
It was then used to project 2002 traffic and as alluded to above, it
overestimated traffic by 3.8%, in a year where traffic hardly grew. Using the traffic per user concept of
projecting future network loads worked regardless of growth or no growth
conditions. The method presented here
provides an easy way to predict future traffic loads based on factors that
should be readily available at any company.
Don’ let the formulas
below scare you off, you did harder math in your high school algebra class.
The Method
Calculating 'Average Traffic
per User', ∆y
Dividing past traffic (T)
by headcount (HC) produces an 'average
traffic per user' (ATU), this number is
calculated monthly.
ATU = T / HC
Each month the change from the previous month is
calculated as a ‘delta of growth’ (∆)
of the average user generated traffic.
∆ = (
ATUC - ATUC-1 ) / ATU C-1
Where ATUC
is the current month's average traffic per user and ATUc-1
is the previous month's average traffic per user. The result, or delta of growth (∆), is the percent of change in the traffic
generated by 'an average user' from the previous month.
The previous year of these deltas are averaged,
∆y = ∑ {∆JAN,
∆FEB, ∆MAR, …, ∆DEC} /
12
The sum of the monthly deltas is divided by 12
yielding the average monthly growth for the year, ∆y,
for an average user. ∆y is the percentage factor that
will be applied to projected headcount data to generate future traffic
projections. Of course the same process
could be done for any other period of interest, such as three or six months.
Headcount Projection Presumption
In order to project traffic based on headcount some presumptions must be made about future employment numbers. Typically HR’s will not provide projected employment information, but they do release periodic general announcements about future layoffs and hiring's. In most cases this will be the only information available for projecting headcount. With nothing else to work with, just presume no change in headcount.
Calculating 'Traffic
Projection', TP
Given a projected headcount (HCp), the average traffic per user per month (ATUC) and the projected yearly growth
rate (∆y ) the next months
traffic can now be calculated (ATUp). Using the Current months Average Traffic per
User multiplied by the growth factor percentage plus one yields the Projected
Average Traffic per User, (ATUp)
for the next month.
ATUp =ATUC * ( 1 + ∆y
)
Multiplying this result by the projected number of
employees (HCP) gives the
projected traffic count (TP).
TP
=ATUp * HCp
This process is repeated for each successive month to be projected.
TP+1
=ATUp+1 * HCp+1
That is all that is required to
project future traffic, but if you want to try and add some additional reality
factors into your calculations use the following method.
Dealing with recurring Anomalies
The December Drop
A pattern has been noted at my company that every
December there is a noticeable drop in the amount of traffic generated
enterprise wide, even though the data is normalized each month to account for
the difference in the number of days of the month[1]. This drop has been much more pronounced in
the last five years, so those are the years used to calculate future December
drops in traffic levels
∆DD
=∑ {∆DEC1997, ∆DEC1998, …, ∆DEC2001} / 5
The
factor (∆DD) is created by
averaging the Monthly Percent of Change of the last five Decembers. The ∆DD
percentage is then applied to future Decembers in any projection.
In our case this phenomena is always in December and
only lasts one month, as you become familiar with your data you may notice
recurring patterns. If you can quantify them add them in.
The January
Adjustment
As a result of the annual December Drop the January
projections cannot be based on the previous months data as in all other cases. The projections for each January’s growth are
based on the previous November’s data rather than on the December results.
Example Calculation
Calculating 'Average Traffic
per User', ∆y
Average traffic per user per month
calculation, using hypothetical numbers,
T = 200 Tb, Total traffic for the month
HC = 200000,
Total headcount for the month
ATUC = T / HC
= 200 Tb of traffic / 200000 people
= 1 Mb per user per month
Change in traffic per user is
calculated
ATUC-1
= .97 Mb, last months traffic per user
∆ =( ATUC - ATUC-1
) / ATU C-1
∆ = ( 1 Mb - .97 Mb ) / .97 Mb
∆ = 0.03093
Meaning there was 3.1% growth from one
month to the next
The previous year of these deltas are averaged,
∆y =∑ {∆JAN, ∆FEB, ∆MAR,
…, ∆DEC} / 12
∆y = (.002 + .034 + .020 + .021 + .022 + .023 + .025 +
.034 + .020 + .021 + .022 + .031) / 12
∆y = 0.02291
The average growth per user per month
was 2.3% over this period.
Calculating 'Traffic
Projection', TP
Calculating next months traffic per
user,
ATUC
= 1 Mb, current average traffic per user
ATUp =ATUC * ( 1 +
∆y )
ATUp = 1 Mb * ( 1 +
0.031 )
ATUp = 1.031 Mb
Next months expected traffic per user
is 1.031 Mb
Calculating next months total
traffic,
HCp
= 200020, number of people expected to be employed next month
TP =ATUp * HCp
TP = 1.031 Mb * 200,020
TP = 206220.62 Mb
Next months traffic is projected to be
206 Tb
Continuing with the next month’s calculations
building off last month’s projections,
HCp+1
= 200010, number of people expected to be employed next month
ATUP+1 = 1.031 Mb * (1 +
0.031)
ATUP+1 = 1.063 Mb
TP+1 =ATUp+1 *
HCp+1
TP+1 = 1.063 Mb * 200,010
TP+1 = 212611 Mb
The following months traffic is
projected to be 212 Tb
This process is repeated for each successive month
to be projected.
Dealing with recurring Anomalies
Lets say you find that every month in summer your
traffic per user drops by 15% from average.
You can apply that factor to the projected traffic per user to refine
your results. Below the monthly traffic
total is recalculating with the summer drop-off factored in.
ATUP+1 = 1.031 Mb * (1 + 0.031
- 0.15)
ATUP+1 = 0.908 Mb
TP+1 =ATUp+1 *
HCp+1
TP+1 = 0.908 Mb * 200,010
TP+1 = 181609 Mb
The adjusted traffic total is now 181
Tb
As you can see from this example, if you can spot recurring changes in traffic, and can quantify them, they can have a big affect on future projections.
Lessons of Experience
We are now starting to apply this method to smaller
sets of data, such as regional areas, specific sites and/or links. Naturally the old axiom of garbage in,
garbage out still holds. The more
accurate the data you use to start with the closer the projections can come to
reality.
If you can get the number of actual users, not full
employment; it will increase your accuracy.
You can get some pretty interesting results with
wildly fluctuating data. What I have
found is that by looking for patterns in the data and adjusting the time used
to determine the future growth rate can help bring things back to reality. For instance, if you have monthly data
results over a year, but the first six months of data is very different than
the last six months you might want to only use the last five months to
determine your growth rate for the next year.
As Winston Churchill said, “I do not believe any
statistics until I have altered them myself.” Always remember to ask the
question, “Does that make sense?” If it
doesn’t, re-check your calculations and adjust as required.
Conclusion
This method changed the way that traffic is being
projected at my company, it also changed the way we think about where traffic
comes from. In the past only one variable
was used to predict future traffic, and that variable was past traffic. This new method introduces the idea of adding
a second variable to the projection equation, the number of users generating
traffic, or the traffic per user. Time has shown that this method works at my
company. It is very easy to apply if you
have the data. If you need to get a
handle on what your future traffic loads might be this is an extremely cost
efficient way to get your answers.
This mathematical model has the same hole in it as
all other methods. In reality there is a
known third variable that encompasses all the un-quantifiable changes, spikes,
additions and deletions that a network constantly goes through. It is highly unlikely that this complex
variable can ever be added to the equation, but it must always be considered
when predicting future traffic.
I will send anyone asking for it a copy of an
example Excel file via e-mail.
Steve
McGourty
Bio
Steve
has been working at Boeing for 17 years, all in networking. His first few years were spent evangelizing
the wonders and benefits of networks, the next few as lead LAN designer for an
ever-expanding network in the largest building in the world and beyond. His work for the last 10 years has been
dedicated to Network Management. He has
a B.S. in Systems Science (Scientific Option) from the