The Origins and Future of Distributed Computing and Clouds - The Practice of Cloud System Administration - page 452

if the service is available on each of the machines. Load balancers can be configured in a
primary/secondary setup—sending all traffic to the primary device until it fails, and then
automaticallyswitchingtothesecondarydevice.Theycanalsobeconfiguredtoload-share
between machines that all provide the same service. In this latter mode, when a machine
stops responding correctly, the load balancer stops directing queries to it until it resumes
respondingcorrectly.Loadbalancersprovideautomaticfailoverincaseamachineisdown,
making them a useful tool for a service that needs high availability.
Scaling
Businesses needed servers to power their web sites. System administrators applied their
old methods to the new requirements: one machine for each web site. As the site got more
popular, larger machines were used to meet the demand. To achieve better performance,
custom CPUs and new internal architectures were developed, but these machines were ex-
pensive. Software was also expensive. A typical web server required an OS license, a web
server license, and a database server license. Each was priced proportionately to the hard-
ware.
With the web the requirements for scaling are not bound to the number of employees
in the company. The web introduces an environment where the users of the service can
be anyone with an Internet connection. That is a very large and rapidly growing number.
When a site introduces or successfully markets a web-based service, the number of people
accessing the servers can increase very rapidly over a short period of time. Scaling for
sudden, unpredictable fluctuations in service usage was a new challenge that was not well
understood in this era, although it was something for which Internet startups tried to pre-
pare.Internetstartupsplannedforsuccessandpurchasedthebiggest,mostreliablesystems
available. These were expensive times.
However,thepracticeofgrowingwebsitesthroughbigger,morereliable,moreexpens-
ive hardware was not economically viable. Normally as a company grows, economies of
scale result in lower cost per unit. The dot-coms, however, required computers that were
more expensive per unit of capacity as the company grew larger. A computer that is 10
times more powerful than an average computer is more than 10 times as expensive. These
larger,morereliablesystemsusedcustomhardwareandhadasmallermarket—twofactors
that drove up prices. For linear increases in performance, the cost per unit of capacity was
growing super-linearly. The more users the web site gained, the more expensive it was to
provide the service. This model is the opposite of what you want. Additional costs came
from the techniques used to ensure high availability, which are covered in the next section.
Oddly enough, these high costs were acceptable at the time. The dot-coms were flush
withcash.Spendingitonexpensivemachineswascommon,becauseitshowedhowoptim-
istic you were about your company's success. Also, the economics of scaling sites in this