What the BYTE Benchmarks Test

The Native Mode benchmarks represent a collection of 10 diverse algorithms. With the exception of the Numeric Sort and String Sort tests (which both use the same algorithm, but not the same code), each test is significantly different from the others. Hence, the system being benchmarked is exposed to a variety of code profiles: There are sorts, searches, matrix operations, compressions, encryptions, and more. Here is a description of the 10 benchmark tests.

Numeric Sort . The Numeric Sort benchmark measures the time it takes to sort a one-dimensional array of signed long integers. It's built around the well-known heapsort algorithm. This is a good, general-purpose test of processor horsepower, since sorti ng is a fundamental operation that's found inside applications ranging from databases to word processors to operating systems.

String Sort . Like the Numeric Sort, the String Sort benchmark has at its heart a heapsort algorithm. However, this test juggles strings of bytes rather than fixed 32-bit integers, thereby putting pressure on the system's ability to move arbitrarily long blocks of bytes to and from arbitrary address boundaries (something, for example, that word processors must often do).

Bitfield . The Bitfield benchmark exercises a system's ability to manipulate single bits. The test is actually constructed as a type of simulation. Specifically, the test mimics what might happen inside an operating system that uses a bit map in memory to keep track of the allocation of disk blocks.

Emulated Floating-Point . This test is fairly self-explanatory; it performs fundamental math operations--addition, subtraction, multiplication, and division --with an IEEE-compliant floating-point package that makes no use of the math coprocessor. (Although the number format that we use is not strictly IEEE-compatible, it would be trivial to write translation routines to convert to and from true IEEE numbers.)

Fourier Coefficients . This benchmark calculates the first n Fourier coefficients for a cyclic waveform constructed using a logarithmic function. This algorithm exercises a system's trigonometric functions.

Assignment Algorithm . The Assignment Algorithm benchmark has a direct application to the business world. Basically, it solves a simulated resource-allocation problem, and in doing so it performs a variety of operations on two-dimensional integer arrays.

Huffman Compression . This benchmark executes the well-known Huffman-method compression algorithm, which is still in use, in one form or another, within some graphics file formats. The routine combines text processing, management of comple x data structures (i.e., the benchmark constructs a kind of binary tree in memory), and bit-manipulation operations.

IDEA Encryption . The IDEA (International Data Encryption Algorithm) is a relatively new and powerful algorithm for encrypting digital data. IDEA is a block cipher that operates on a group of 16 bits at a time. The benchmark test measures how quickly a system can encrypt and decrypt a byte stream.

Neural Net . The Neural Net benchmark test is based on a simple back-propagation neural network, as presented by Maureen Caudill in her article ``Expert Networks'' (October 1991 BYTE). The neural net is taught to recognize a number of ASCII characters. The resulting test is primarily a floating-point benchmark that makes heavy use of the exponential function.

LU Decomposition . The LU Decomposition benchmark is constructed around an algorithm of the same name that can be used to--among other things--solve systems of linear equations. Th is benchmark primarily measures a system's fundamental floating-point capabilities: addition, subtraction, multiplication, and division.