Elapsed time and throughput are two different ways of measuring speed. Elapsed time is the better measure for processor speed because it is less dependent on other system components. When processor performance is quantified, it is taken to be inversely proportional to execution time.
There are two different ways of dealing with processor speed. Someone who is purchasing a computer is primarily concerned with the "bottom line" - how fast it will do whatever the customer wants to do with it. The assessment needs of customers are best met by benchmarks.
Computer architects cannot rely on just performance assessment. Assessment requires an implementation of a processor, so it is of little use in designing a new processor or making modifications to an existing processor. Computer architects need performance analysis tools to enable them to predict performance improvements.
When measuring how fast something is, there are two kinds of measures. You can measure how long it takes to do something or you can measure how much gets done per unit time. The former is referred to as response time, access time, transmission time, or execution time depending on the context. The latter is referred to as throughput.
When measuring how fast something is, there are two kinds of measures. You can measure how long it takes to do something or you can measure how much gets done per unit time. The former is referred to as response time, access time, transmission time, or execution time depending on the context. The latter is referred to as throughput.
Elapsed Time ThroughputElapsed time speed measures measure the elapsed time from the initiation of some activity until its completion. The phrase response time is often used in operating systems and graphical user interface contexts. The phrase access time is used for evaluating data storage systems. The phrase nodal delay is used for evaluating links in computer networks. The phrase execution time is used for evaluating processors.
A throughput measure is an amount of something per unit time. For operating systems throughput is often measured as tasks or transactions per unit time. For storage systems or networks throughput is measured as bytes or bits per unit time. For processors, the number of instructions executed per unit time is an important component of performance.
When processor performance is quantified, it is taken to be inversely proportional to execution time. Consequently, performance ratios are inverted from time ratios. That is, a performance ratio is computed as follows.
| performance improvement ratio | = | performance after change | = | execution time before change | ||||||
| performance before change | execution time after change |
Someone who is purchasing a computer is primarily concerned with the "bottom line" - how fast it will do whatever the customer wants to do with it. The customer usually has little interest or technical skill for understanding the underlying causal factors.
The assessment needs of customers are best met by benchmarks. The Standard Performance Evaluation Corporation (SPEC) specifies two current suites of execution time benchmark programs:
These suites are widely used today for comparing the performance of processors. The programs are also useful for providing information for predicting performance after incremental CPU design changes.
Computer architects cannot rely on just performance assessment. Assessment requires an implementation of a processor, so it is of little use in designing a new processor or making modifications to an existing processor.
Like software developers, computer architects spend more time enhancing existing processors rather than developing new processors. Because of this, they are often interested in comparing an enhanced processor version to the current version. These comparisons can be expressed either as improvement ratios or improvement percentages.
Architects need to be able to break down performance into components that are easier to understand and predict. The components can help architects understand where obstacles to good performance lie, identify their causes, and predict the effects of modifications.
Architects use two important analytical tools for dealing with performance.