Dhrystone 2.0 timings

[David L. Luner]

In a previous article shawn at jdyx.UUCP (Shawn Hayes) writes:

>      Well, this want answer all the questions about how fast the new RS 6000
>machines really are, but here is my attempt to help. 

Thank you. All we're (and, I suppose, We're all) looking for is a fair 
evaluation.

>     I recently got a copy of the Dhrystone program V2.0 and ran it on our
>RS 320 machine.  

The current field releases of AIX V3.1 are preliminary. Furthermore,
they are changing daily. I believe the current release is "8943Q". What
version we you running on? If you run benchmarks on early releases, be
aware that things may change *in either direction*.

> The results aren't quite what IBM stated (using V1.0), but
> that shouldn't be a big suprise.  :>  

For any number of reasons. First: they even ran the Dhrystone V1.1 on
a different release. Second: Dhrystone V2.0 makes things (intentionally)
more difficult for the compiler. Third: Read on for caveats w.r.t. "43Q".

> I got a value of 11,881.2 dhrystones/sec
>using a non-optimized version, and a value of 18,867.9 dhrystones/second
>using the optimizer.  Averages out to about 11 MIPS which isn't bad.  

In trying to duplicate the published Dhrystone V1.1 numbers that were
published I discovered two things on. For the time being I only assume
this is specific to "43Q". Number one, the clock runs at 100Hz. The
V1.1 code I was using assumed a 60Hz clock so the numbers looked low.
Secondly, there is a problem in the C libraries related to the string
handling functions. A bypass (to convince you of the "reasonableness"
of the published numbers) is to

	#include <strings.h>

This will in-line the string handling functions. *** I know this is not
kosher. Until the C library is fixed in a later release, it will have to
do.

>  
>     In addition I decided to run another test today and ran 11 copies of the
>optimized version at the same time.  I got about 18,700 dhrystones/sec on
>10 of the copies and about 14,000 on the last copy.  Not bad at all.
>

Assuming that the Dhrystone V2.0 code is like the V1.1 code in this
respect, it only looked at user process time, not elapsed clock time.
If I am wrong in this, I am impressed because what you stated implies
over 187K Dhrystones.

>    Speaking of benchmarks I have two requests.  
>First, does anyone know of any good system level benchmarks?  

IBM has stated that SPEC (for sure) and RAMP-C (I think) numbers will be
provided in the forseeable future. Other things you may look at are the
MUSBUS and X-benchmarks.

> Second,  does anyone know how much time an RS 6000/320 will take to do a 
> context switch??????   We've talked to our SE...

Please refer your SE to me.


	-- David

David Luner
IBM Madison
(608) 273-5243

Disclaimer: #include <sys/aix/rs6000/disclaimer>