home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Nebula 1
/
Nebula One.iso
/
Utilities
/
BenchMarks
/
ByteBenchmark
/
doc
/
bench3.doc
< prev
Wrap
Text File
|
1994-01-27
|
2KB
|
47 lines
The BYTE Unix benchmarks that we published in March, 1990, we
expected that they fill our needs for several years. It was only a
few months before we started to see that UNIX workstation performance
was going to make the benchmark design questionable. The design flaw
was that the benchmarks timed a fixed number of loops; if there were
to few loops, the times were too small to be reliable. Perhaps we
could have increased the number of loops and been safe for another
few years (months?). But with IBM's introduction of the RS/6000, it
became obvious that we needed to redesign our benchmarks.
Now the work on version 3 is complete. The new design increases in
precision on faster workstations because we turned everything
upside-down: we count loops for a fixed amount of time. There are
some additional benefits, the amount of time for running the suite is
relative predictable (about an hour).
Since new benchmarks mean new baselines and indexes, we reevaluated
these elements as well. Though the Everex Step 386/33 is still a very
contemporary machine, the operating system that we were using for the
baseline (SCO Xenix 386 version 2.3.1). We were tempted to use the
new HP 9000 Series 700 as the baseline, but so few machines are in
that league that the index would be very strange. Instead, we have
taken the conservative approach: a popular machine. The new baseline
machine is the Sun Microsystems [TOM, please fill in configuration],
the baseline for this month's roundup of SPARC clones.
We also changed which benchmark tests are used to generate the new
index:
* a double precision arithmetic performance test
* Dhrystone 2 without register variables
* Spawning a process (!M!execl()!EM!)
* File copy throughput in five seconds
* Pipe-based context switching
* The shell script with 8 concurrent scripts running
Instead of just summing the indexes of each of these tests to get the
overall index, we are using the average. This means that the baseline
value is 1 instead of six.
As always, the BYTE Unix benchmarks are freely available from BIX and
on the Usenet. Version 3 automatically configures to System V and BSD
Unix systems. It also includes scripts for generating the indexes.
Thanks to the many users who sent use reports and suggestions for
Version 2, BYTE and you have a much better set of tools for
benchmarking UNIX systems.