You are here

Clock rate does not equal performance!

1 June, 2015 - 16:42

Computer vendors often use performance numbers to indicate the advantages of their systems and to compare against their rivals. Those performance numbers are usually obtained from highly-tuned programs with tests that focus on highly-isolated criteria, such as the raw ability of a processor to perform a specific set of calculations.

However, the fact is that the performance improvement of a computer system for general applications is not necessarily proportionate to the increase in speed of the processor. Let's take the systems built with Intel Pentium 4 processors as an example -- the MPEG digital video encoding performance only improved by 23% with a 42% increase in the processor speed on the Intel Pentium 4 systems.

Computer systems today are built with tens of thousands of individual components that are interconnected together. Each component performs its duty independently when a request is received from its interconnected neighbours. Unless the performance of every single component in the computer system improves at the same rate, the overall improvement experienced by the applications will be limited.

Let's revisit the story of Mr ProudOfIt and Mr NoClue.

Since networking and disk I/O activities are individual components that operate independently of the computer's processor, an increase in processor speed will have little or no effect on the performance of the networking and disk I/O activities. Unless the performance of the networking and I/O components are improved at the same rate as the system processor, we cannot expect a 6X improvement in the surfing experience for Mr ProudOfIt and Mr NoClue. This phenomenon can be observed through Amdahl's Law , a concept that we will formalize later in this unit.

Amdahl's Law

Gene Amdahl, a pioneer in our field of study, stated the following:

If F is the fraction of a calculation that is sequential, and (1-F) is the fraction that can be parallelized, then the maximum speedup that can be achieved by using P processors is 1/(F+(1-F)/P). (Amdahl, 1967)

This is called Amdahl's Law, one of the fundamental laws in studying computer system performance. It states that the performance improvement to be gained from using a faster mode of execution is fundamentally limited by the fraction of the time the faster mode can be used.]