PERFORMANCE 551

tions, particularly those that purchase or use many machines extensively, have one or more programs that they believe characterize their own workload. Whether a standard benchmark can be of value in characterizing performance depends on the degree that it is typical of the actual use of the computer. A further advantage of benchmarks is that they are the language that the computer is to use, and, hence, reflect the application and characterize the language machine architecture. To illustrate the variability in the scientific FORTRAN benchmark metrics, the performance of a number of machines (VAX- 11/780 with floating-point accelerator option, PDP- 11/70, and DECSYSTEM 2060), executing about a dozen such benchmarks, is compared in Figure 2. Two scientific benchmarks of the National Physical Laboratory in the United Kingdom [Curnow and Wichmann, 1976] are often singled out as being the most useful benchmarks because of the extensive effort that was put into designing them as typical scientific programs. Several factors, such as the frequencies of the trigonometric functions, frequencies of subroutine calls, and characteristics of the I/O, were considered. The performance of computers executing these benchmarks is expressed in Whetstones per second.

There are similar benchmarks for commercial processors that generally use the COBOL language.