Monitoring Architecture and Practice - The Practice of Cloud System Administration - page 357

ure. A network link that is overloaded during the day and barely used at night will be, on
average, half used. It would be a bad idea to make a decision based on the average utiliza-
tion.Averageslie.IfbillionaireBillGateswalksintoahomelessshelter,theaverageperson
in the building is a multimillionaire but it doesn't take a genius to realize that the homeless
problem hasn't been eliminated.
Thatsaid,averagesaren'tinherentlymisleadingandhaveaplaceinyourstatisticaltool-
box. Like other statistical functions they need to be used wisely.
17.5.1 Percentiles
Percentiles and medians are useful when analyzing utilization or quality. These terms are
often misunderstood. Here's a simple way to understand them.
Imagine a school has 400 students. Stand those students up along a 100-meter line in
orderoflowest GPAtohighest.Stretch outthelineevenlysothatthestudentwiththelow-
est GPA is at 0 and the one with the highest GPA is at 100. The student standing at the
90-meter mark is the 90th percentile. This means that 90 percent of the student body did
worse than this person. The student at the 50-meter mark, exactly in the middle, is known
as the median.
Iftherearemorestudents,iftherearefewerstudents,orifthegradingstandardchanged,
there will always be a student at (or closest to) the 90-meter mark. The GPA of the 90th
percentilestudentcouldbeaC-oranA+.Italldependsonthemakeupoftheentirestudent
body's GPAs. The entire school could have excellent grades but someone is still going to
be at the 90th percentile. Trends can be analyzed by, for example, graphing the GPA of the
50th and 90th percentile students each year.
It is common for colocation facilities to charge based on the 90th percentile of band-
width used. This means they take measurements every 5 minutes of how much bandwidth
is being used at that moment. All these measurements are sorted, but duplicates are not re-
moved.Ifyouwroteallthesemeasurements,distributedevenly,alonga100-meterfootball
field,themeasurementatthe90-metermarkwouldbethebandwidthrateyouwillbebilled
for. This is a more fair billing method than charging for the maximum or average rate.
It is very smart of your colocation facility to charge you this way. If it charged based on
the average bandwidth used, the company would take the total amount ofbytes transmitted
in the day, divide by the number of seconds in a day, and use that rate. That rate would be
misleadinglylowforcustomersthatconsumeverylittlebandwidthatnightandalotduring
the day, what is known as a diurnal usage pattern. In contrast, the 90th percentile repres-
ents how heavily you use bandwidth without penalizing you for huge bursts that happen
from time to time. In fact, it forgives you for your 30 most extreme 5-minute periods of