Re: performance measurement and caches

Hans Boehm wrote:
>
> An anecdote from a few days ago:
> [benchmarks on a SPARCstation 10 differed in a factor of 2 depending on
> where in the cache they landed]


A Dos box is actually very nice when you want to get repeatable
numbers, both because there is very nearly no background traffic,
except for the 18.2 Hz clock tick, and because it is easy to set up
the timing to start after the program is loaded, or load it from a ram
disk.


When timing som fp code for a Pentium vs PentiumPro comparison, I did
repeated sets of micro-measurements, using the RDTSC opcode to get the
cpu cycle counts for each piece of code.


Before the timing part, I did a variable (but controlled) amount of
cache flushing, by reading sequentially through a memory buffer, from
zero to 256K in size.


This way I could repeat each measurement 120 times, and get exactly
the same cycle count on more than 110 of them.


--
-Terje Mathisen (include std disclaimer) <Terje.Mathisen@hda.hydro.com>
[Doesn't help a SPARC much, but cycle counters might indeed be useful. -John]
--

Return to the comp.compilers page.
Search the comp.compilers archives again.