minix/benchmarks/unixbench-5.1.2/WRITING_TESTS
2010-02-03 14:30:02 +00:00

133 lines
4.5 KiB
Text

Writing a Test
==============
Writing a test program is pretty easy. Basically, a test is configured via
a monster array in the Run script, which specifics (among other things) the
program to execute and the parameters to pass it.
The test itself is simply a program which is given the optional parameters
on the command line, and produces logging data on stdout and its results on
stderr.
============================================================================
Test Configuration
==================
In Run, all tests are named in the "$testList" array. This names the
individual tests, and also sets up aliases for groups of tests, eg. "index".
The test specifications are in the "$testParams" array. This contains the
details of each individual test as a hash. The fields in the hash are:
* "logmsg": the full name to display for this test.
* "cat": the category this test belongs to; must be configured
in $testCats.
* "prog": the name of the program to execute; defaults to the name of
the benchmark.
* "repeat": number of passes to run; either 'short' (the default),
'long', or 'single'. For 'short' and 'long', the actual numbers of
passes are given by $shortIterCount and $longIterCount, which are
configured at the top of the script or by the "-i" flag. 'single'
means just run one pass; this should be used for test which do their
own multi-pass handling internally.
* "stdout": non-0 to add the test's stdout to the log file; defaults to 1.
Set to 0 for tests that are too wordy.
* "stdin": name of a file to send to the program's stdin; default null.
* "options": options to be put on the program's command line; default null.
============================================================================
Output Format
=============
The results on stderr take the form of a line header and fields, separated
by "|" characters. A result line can be one of:
COUNT|score|timebase|label
TIME|seconds
ERROR|message
Any other text on stderr is treated as if it were:
ERROR|text
Any output to stdout is placed in a log file, and can be used for debugging.
COUNT
-----
The COUNT line is the line used to report a test score.
* "score" is the result, typically the number of loops performed during
the run
* "timebase" is the time base used for the final report to the user. A
value of 1 reports the score as is; a value of 60, for example, divides
the time taken by 60 to get loops per minute. Atimebase of zero indicates
that the score is already a rate, ie. a count of things per second.
* "label" is the label to use for the score; like "lps" (loops per
second), etc.
TIME
----
The TIME line is optionally used to report the time taken. The Run script
normally measures this, but if your test has signifant overhead outside the
actual test loop, you should use TIME to report the time taken for the actual
test. The argument is the time in seconds in floating-point.
ERROR
-----
The argument is an error message; this will abort the benchmarking run and
display the message.
Any output to stderr which is not a formatted line will be treated as an
error message, so use of ERROR is optional.
============================================================================
Test Examples
=============
Iteration Count
---------------
The simplest thing is to count the number of loops executed in a given time;
see eg. arith.c. The utlilty functions in timeit.c can be used to implement
the fixed time interval, which is generally passed in on the command line.
The result is reported simply as the number of iterations completed:
fprintf(stderr,"COUNT|%lu|1|lps\n", iterations);
The bnenchmark framework will measure the time taken itself. If the test
code has significant overhead (eg. a "pump-priming" pass), then you should
explicitly report the time taken for the test by adding a line like this:
fprintf(stderr, "TIME|%.1f\n", seconds);
If you want results reported as loops per minute, then set timebase to 60:
fprintf(stderr,"COUNT|%lu|60|lpm\n", iterations);
Note that this only affects the final report; all times passed to or
from the test are still in seconds.
Rate
----
The other technique is to calculate the rate (things per second) in the test,
and report that directly. To do this, just set timebase to 0:
fprintf(stderr, "COUNT|%ld|0|KBps\n", kbytes_per_sec);
Again, you can use TIME to explicitly report the time taken:
fprintf(stderr, "TIME|%.1f\n", end - start);
but this isn't so important since you've already calculated the rate.