Databases Reference
In-Depth Information
would include the processing power of all human minds on the planet. The conse-
quence for database designers (and most scientific computing and processing intense
applications) is that the processing power you purchase for your server at any moment
in time will usually be supplanted in less than two years by technology that is approxi-
mately twice as fast. See Figure 13.3.
13.1.2 Amdahl's Law for System Speedup with Parallel Processing
In principle more processors should mean more processing capability, and therefore
faster execution time. In practice several factors conspire—in logistics, physics, and eco-
nomics—to prevent this from being possible. The most obvious is the simple observa-
tion that not all tasks are well suited to parallel processing. Three women cannot make a
baby in three months. Owning three cars still does not make it possible to drive from
Boston to San Jose in one-third of the time.
A simple expression of the maximum potential benefit from parallel processing was
defined by Amdahl's Law.
2
Amdahl's Law considers the optimal potential processing
gain assuming overhead processing is nil. To do this the model separates the tasks that
can be performed in parallel from the tasks that need to be serialized. The speedup time
on
N
processing,
S
(
N
), is defined as the ratio of execution time on a single processor
T
(1) to the execution time on
N
processors
T
(
N
). Amdahl subdivides the time
T
into
serial processing time
Ts
and parallel processing time
Tp
, and then expresses the optimal
speedup as:
13.1
However, even tasks that are well suited to parallel processing often do not improve
in performance with additional CPUs. The major limitation is generally called “over-
head,” and includes factors such as context switching, setup time (stack operators,
thread or process initialization), bus contention, etc. When the overhead time is large
compared to the processing time one can expect parallel processing to perform poorly
and in some cases even cause performance to get worse. This overhead is often nonlin-
ear, for example, the cost of bus contention, and context switching can increase in non-
2
Gene Myron Amdahl is best known for his work in mainframe computers at IBM in the 1950s
and 1960s. He founded Amdahl Corporation in 1970; manufactured “plug-compatible” main-
frames, shipping its first machine in 1975. The 1975 the Amdahl 470 V6 was a less-expensive,
more reliable, and faster computer than IBM's System 370/165. By 1979 Amdahl had more than
6,000 employees and had sold more than $1 billion U.S. in mainframe equipment. Gene left
Amdahl in 1979 and went on to found and consult for numerous other technology firms.
