Page 241 -

Chapter 6 The Cloud
240
0GGF 5ECNCDKNKV[

2GCM WUCIG QXGT )DRU

8QNWOG QH &CVC &GNKXGTGF )DRU

Figure 6-1 #/ #/ 2/ 2/ 2/ 2/ 2/ 2/ 2/ 2/ 2/ 2/

Example Video Banner #/ #/ 'ZCORNG 8KFGQ DCPPGT CF EWUVQOGT
Ad Customer

ad may appeal more to one nationality than to another. Will 70 percent of those visits arise in the
United States and the rest in Europe? Or will there be millions from Japan? Or Australia? Given this
uncertainty, how does the car manufacturer prepare its computing infrastructure? The car manu-
facturer knows that if it cannot provide very short response time (say, a fraction of a second), it
will lose the benefit of an incredibly expensive ad. On the other hand, if the ad is a flop, preprovi-
sioning of thousands of servers will add to the accumulation of wasted money.
Figure 6-1 shows an example of this situation, based on a real case supported by Amazon.
com’s CloudFront. Suppose Figure 6-1 shows the processing on the car manufacturer’s Web site
during the Academy Awards. Throughout the day, the car manufacturer is delivering less than
10 Gbps of its content to users. However, as soon as its ad runs (2 PM in the Hawaii-Aleutian
time zone where the data was collected), demand increases seven-fold and stays high for half an
hour. After the announcement of Best Picture, when its ad runs again, demand again increases to
30 and 40 Gpbs for an hour and then returns to its base level.
Without an increase in servers, response time will be 3 or 5 seconds or more, which is far too
long to maintain the attention of a charged-up Academy Awards viewer. However, the car manu-
facturer has contracted with its cloud vendor to add servers, wherever needed worldwide, to keep
response time to less than 0.5 seconds. Using cloud technology, the cloud vendor will program-
matically increase its servers to keep response time below the 0.5-second threshold. As demand
falls after the ad runs a second time, it will release the excess servers and reallocate them at the
end of the awards.
In this way, the car manufacturer need not build or contract for infrastructure that supports
maximum demand. Had it done so, the vast majority of its servers would have been idle for most
of the evening. And, as you’ll learn, the cloud vendor can provision servers worldwide using the
cloud; if a good portion of the excess demand is in Singapore, for example, it can provision extra
servers in Asia and reduce wait time due to global transmission delays.

Pooled
The second key in the definition of cloud is pooled. Cloud resources are pooled because many
different organizations use the same physical hardware; they share that hardware through
virtualization. Cloud vendors dynamically allocate virtual machines to physical hardware as cus-
tomer needs increase or decrease. Thus, servers that advertisers need for the Academy Awards
can be reallocated to CPA firms that need them later that same day, to textbook publishers who
need them for online student activity on Monday, or to the hotel industry that needs them later
the next week.
An easy way to understand the essence of this development is to consider electrical power.
In the very earliest days of electric power generation, organizations operated their own genera-
tors to create power for their company’s needs. Over time, as the power grid expanded, it became