Programmer 7500/MAX_PROGRAMMERS.iso/PROGRAMS/UTILS/BENCHMAR/MCHALL10.ZIP/MCHALL10.DOC

Text File | 1990-11-12 | 5.4 KB | 126 lines

                         THE MCHALL BENCHMARK
                             VERSION 1.0
                               11/12/90



1. BACKGROUND

         In comparing the performance capabilities of computers as
possible platforms for handling signal processing requirements, the
existing benchmarks (Whetstones, Drystones, MIPS, etc.) are often
confusing and contradictory.  In addition, they may measure parameters
that are of no benefit in the application being considered.

         The MCHALL Benchmark was conceived as a portable, quick-and-dirty
test that measures computer performance in accomplishing a narrow range of
tasks.  The tasks were selected to represent typical processing
requirements.  The benchmark tests include (1) reading a block of data
from a file, and (2) performing a series of arithmetic operations on that
block of data.

         The program was deliberately kept short and simple, and was
implemented without prejudice toward or against any computer.  It is felt
that detailed comparisons between computers running actual application
programs may differ from the MCHALL results; nevertheless, the MCHALL
Benchmark should provide a first level indication of relative performance.

2. IMPLEMENTATION AND EXECUTION

         For reasons of portability, the MCHALL Benchmark was implemented
in the ANSI C language.  Most computers are provided with C compilers that
run ANSI C source code, therefore, apples versus apples testing can be
maintained.

         When the MCHALL test is run, the operator is asked to enter the
desired number of loops to be executed.  One loop consists of:

         --  Reading a block of 10,000 16-bit integer values
         from a file (MCHALL.DAT) that will have been previously
         created by the MCHALL program.

         --  Executing a loop for 50,000 cycles.  The loop
         contains the following instruction types:

                 15  Additions and subtractions
                  1  Multiplication
                  3  Divisions
                  1  Shift
                 10  Comparisons
                 10  Branch Decisions

         In very slow machines, the time required to run one loop may be
sufficient to provide a very low timing error tolerance -- in fast
machines, multiple loops must be run to reduce the error tolerance.  (The
C timing instructions used provide an overall timing error of plus or
minus 0.5 second over the duration of the test.  Therefore, if the machine
is fast, and completes the run in one second, the error tolerance is ±0.5
second or 50%.  This problem is reduced by running the test over a higher
number of loops on faster machines.)

         To maintain a consistent results database, the following "rules"
have been established for running the tests:

         1.  Compiler optimization features (if available) should
         be disabled prior to compilation.  (We want all computers
         to run the same set of instructions for the tests.)

         2.  On computers containing floppy and hard disk drives,
         the Benchmark should be run from the hard drive.

         3.  The user should set the "loop" value high enough to
         achieve an error tolerance of less than ±1%.

         4.  On multi-user computers, no other users are allowed
         to be active during the testing period.

         NOTE:  In order to maximize portability of the source code, the
.DAT data file that is created by the program is not deleted.  Some compilers
were found to have trouble with some of the instructions used to delete
the file.

3.  RESULTS

         As of this writing, the results indicated in the following
tabulation have been obtained using the BETA version of the program.  The
writers would be anxious to entertain constructive suggestions for
modification and further applications of the MCHALL Benchmark.

         Date    Computer                MCHALL  Tol%    Comments

         10/23   PC XT                   133.7   0.12
         10/23   PC XT (Turbo)            80.8   0.12
         10/23   PC 386 (8 MHz)           24.7   0.10
         10/23   PC 386SX (16 MHz)        12.8   0.78
         10/23   PC 386 (20 MHz)           9.5   0.26
         10/23   VAX 11/780                8.9   0.28
         10/24   VAX 3400                  5.8   0.86
         10/24   VAX 3300                  5.7   0.88
         10/24   MAC II CX                 5.5   0.91
         10/24   VAX 3100                  5.4   0.93
         10/24   PC 386 (33 MHz)           5.0   0.50
         10/24   MAC II FX                 2.2   2.27    Tol > 1%
         11/08   PC 486 (33 MHz)           0.99  1.67    Tol > 1%
         10/24   DECstation 5000           0.9   1.11    Tol > 1%


4.  A REQUEST

         This program is contributed to the public domain with the usual
disclaimer (no guarantee of results, useful or otherwise; no liability if
it trashes your machine, etc.).  The authors would appreciate any feedback
on results you obtain using this program, particularly for machines not
listed above.  If someone has access to a CRAY, that would be very
interesting information.

The author of this program is Charles Hall; consultant and beta tester
Thad McCulloch.  Send any feedback to Thad McCulloch at the Exec-PC
bulletin board (414-789-4210), or via PRODIGY (GBNB50A).