> From: Henrik Gedenryd [mailto:Henrik.Gedenryd at lucs.lu.se]
> I've seen tinyBenchmarks slowdowns on the order of 3-5% from 
> merely filing in change sets
[...]

Have you ever seen it speed up?  It wouldn't surprise me if the placement of
frequently-referenced objects relative to the processor cache made a
difference, especially on Intel architectures.  tinyBenchmarks will probably
fit in cache if all the accesses hit different cache lines, but all you need
is for two frequently-accessed addresses to flush each other out of the
cache and cause waits while they're re-loaded.

I'm not a processor expert, and I know there are several on this list.  To
the experts: Could this be the difference?  And can we watch the cache
behaviour on x86 sing the internal monitoring stuff to find out?

		- Peter