[rescue] Sun 711

[James Lockwood]

On Fri, 3 May 2002, George Adkins wrote:

> There was a system reboot prior to the test, anytime I test the speed of
> anything, it's with a clean freshly booted system.

This mitigates most of my earlier concerns.

You should, in general, see close to identical bulk read/write performance
on a nonfragmented fs on a single spindle.

If write performance is noticably slower it indicates a suboptimal
filesystem (or just that Disksuite is a single-threaded pig on writes, but
that's another problem).

> That's nice, however the reason I didn't test with dd in the first place is
> because I wanted to include whatever filesystem overhead (for better or
> worse) there was in the mix.

Per-file overhead percentage depends almost entirely on the size of files
you write.  The reason people use tools like bonnie is that they give you
performance numbers over a mix of access patterns (both sequential and
random, both buffered stdio and read/write) which allow you to extrapolate
behaviour for your specific case.

Bottom line: make sure you're measuring the access pattern(s) you're
concerned with.  This is, of course, the single biggest problem with
benchmarks and benchmarketing.

-James