Information Technology Reference
In-Depth Information
quires less labor. When labor is manual, larger systems require more labor. When opera-
tions are automated, the cost of developing the software is spread across all the machines,
so the cost per machine decreases as the number of machines increases.
To gain further efficiencies of scale, companies like Google and Facebook take this one
step further. Why not treat the entire datacenter as one giant computer? A computer has
memory, disks, and CPUs that are connected by a communication “bus.” A datacenter has
anetworkthatactslikeabusconnectingunitsofcomputeandstorage.Whynotdesignand
operatedatacentersasonelarge,warehouse-sizedcomputer?Barroso,Clidaras,andHölzle
( 2013 ) coined the phrase “warehouse-scale computing” to capture this idea of treating the
entire datacenter as one giant computer.
Costs
In this era companies also competed to develop better and less expensive datacenters. The
open source movement was reducing the cost of software. Moore's law was reducing the
cost of hardware. The remaining costs were electricity and operations itself. Companies
saw that improvements in these areas would give them a competitive edge.
Prior to the second web era, datacenters wasted electricity at unbelievable rates. Ac-
cording to the Uptime Institute, the typical datacenter has an average power usage effect-
iveness(PUE)of2.5.Thismeansthatforevery2.5wattsinattheutilitymeter,only1watt
is used for the IT. By using the most efficient equipment and best practices, most facilit-
ies could achieve 1.6 PUE. The lack of efficiency primarily comes from two places. First,
some efficiency is lost every time there is a power conversion. A UPS converts the power
from alternating current (A/C) to direct current (D/C), and back to A/C—two conversions.
Power is converted from high-voltage lines to 110 VAC used in power outlets to 12 V and
othervoltagelevelsrequiredbythecomponentsandchipsincomputers—fourorfivemore
conversions. In addition, cooling is a major factor. If 1 watt of power makes a computer
generate a certain amount of heat, it takes at least 1 watt of power to remove that heat.
You don't actually “cool” a datacenter; you extract the heat. Companies seeking to reduce
their operational costs like Google, Microsoft, and Facebook achieved closer to 1.2 PUE.
In2011,Googlereportedonedatacenterachieved1.08PUEduringwintermonths( Google
2012 ) . Every decimal improvement in PUE means a big competitive edge.
Other things are even more important than power efficiency. Using less power—for ex-
ample, by shutting off unused machines—is an improvement no matter your PUE rating.
Smart companies work toward achieving the best price per unit of computation. For ex-
ample, for internet search, true efficiency comes from getting the most QPS for your
money. This may mean skipping a certain generation of Intel CPU because, while faster,
the CPU uses a disproportionate amount of power.
Search WWH ::




Custom Search