Branko =?ISO-8859-2?Q?=C8ibej?= <brane@xbc.nu> writes:
>
> It'll be even more obvious to you when I describe how I measured the
> difference. :-)
>
> I modified svndiff-test to leave out the base64 encoder and and the
> actual output of the diff, then tested that on two versions of a 6 MB
> Win32 optimised executable, then measured user time (by the way, the
> un-base64ed svndiff size would've been about 1.5 MB, in 86 windows).
>
> The speedup (using gcc -O2 optimised code!) was 5-15%, depending on
> platform -- and txdelta->svndiff and file reading is included in this
> time, so I suspect the difference for the actual vdelta encoder is even
> larger. Take into account the fact that we calculate the hash -- with
> key length of only 4 bytes -- exactly *once* for every lookup.
>
> Let's just say that I was "mildly" surpriesd by the results ...
May I ask a favor? Could you modify the code to count the number of
collisions, and run that with both hash functions?
I've had weird experiences with benchmarking. I never really feel
comfortable until I've got a concrete explanation for the improvement.
Received on Sat Oct 21 14:36:14 2006