Re: micro-benchmarking

From:

Tom Anderson <twic@urchin.earth.li>

Newsgroups:

comp.lang.java.programmer

Date:

Sat, 2 May 2009 09:45:26 +0100

Message-ID:

<alpine.DEB.1.10.0905020942110.26990@urchin.earth.li>

On Fri, 1 May 2009, Arne Vajh?j wrote:

Mark Space wrote:

Arne Vajh?j wrote:

It does not matter. If milliseconds is not granular enough, then
the results are worthless on OS's like Windows and *nix.

currentTimeMillis(), on Windows at least, could use a very inaccurate timer
that doesn't even compute milliseconds correctly. It's always off by like
18 milliseconds one way or the other (I'd have to look up the exact value).
nanoTime() isn't accurate to nanoseconds, but it's usually accurate to at
least 1 millisecond.

Sure.

But given the unpredictable scheduling of other processes, then you can
still not meaningful measure so small intervals.

Sure you can, you just have to do a lot of runs. This is why when i do
benchmarks, i do 100 runs, and ignore the five slowest, on the grounds
that they probably took so long because either another process was
scheduled in the middle, or GC kicked in or something. For this to be
effective, you actually want to keep the individual runs quite short, to
minimise the chance of them being disrupted, and do a huge number of them,
and for this, a high-resolution timer is essential.

tom

--
I prefer gin now sleep doesn't want me anyway.