SPARC cache effects?

[mr at racal-itd.co.uk]

I'm running some bit crunching code which needs to run as fast as I can
make it. Its been recoded in SPARC assembler (yeuch). Running it on a 4/60
(SunOS 4.1c) I find that different runs on the same data can give a ~15%
difference in real (wall) time, but not usually a time in between. The
machine is otherwise idle. In order to minimise edge effects my test
program reads the data in to memory and then runs the algorithm 5 times.
The 4/60 is not paging. I don't believe that any other process on the
machine is affecting it.

I have a suspicion that I am hitting an artifact of the cache system.  The
code is a few thousand words and the data is typically 100 kbytes.  Any
comments or suggestions welcomed. Please feel free to ask questions.

Regards,
        Martin Reed, Racal Imaging Systems Ltd
uucp: mr at ritd.co.uk, uunet!ukc!ritd!mr
Global String: +44 256 469943   Fax: +44 256 471492
Paper: Rankine Road, Basingstoke, Hants, England, RG24 0NW