Several measur es of perf ormance have been used in the eval uation of compute r systems. The most common ones are: million instru ctions per secon d (MIPS ), million operations per second (MOPS), million floating-point operations per second (MFLOPS or megaflops), billion floating-point operations per second (GFLOPS or gigaflops), and million logical inferences per second (MLIPS). Machines capable of trillion floating-point operations per second (Teraflops) are now available . Lists the common prefixe s used for thes e measures. Po werof-
10 prefixes are typically used for power, frequency, voltage, and computer performance measurements. Power-of-2 prefixes are typically used for memory, file, and register sizes. The measure used depends on the type of operations one is interested in, for theparticular application for which the machine is being evaluated. As such, these measures have to be based on the mix of operations representative of their occurrence in the application. The performance rating could be either the peak rate (i.e., the MIPS rating the CPU cannot exceed) or the more realistic average or sustained rate. In addition, a comparative rating that compares the average rate of the machine to that of other well-known machines is also used. In addition to the performance, other factors considered in evaluating architectures are: generality (how wide is the range ofapplications suited for this architecture), ease of use, and expandability or scalability. One feature that is receiving considerable attention now is the openness of the architecture. The architecture is said to be open if the designers publish the architecture details such that others can easily integrate standard hardware and software systems to it. The other guiding factor in the selection of architecture is the cost. Several analytical techniques are used in estimating the performance. All these techniques are approximations, and as the complexity of the system increases, most of these techniques become unwieldy. A practical method for estimating the performance in such cases is using benchmarks.
