[Zlib-devel] patch-in-progress: vectorized adler32 calculation

[John Bowler]

From: Stefan Fuhrmann
> SVN backend code...
>... reached a point where zlib accounts for almost 40% of the runtime.
> ~15% of the zlib runtime is spent in adler32 

You may find it worthwhile to try tweaking the input and output buffer sizes given to inflate (I assume this *is* inflate).  I found (with an admittedly extreme example) that changing the buffer size could decrease overall execution time by more than 40% (of the original time).

The results are likely to be extremely application dependent and somewhat system dependent, but, given that you are concerned with just one application, it's worth the experiment.  I'd also be interested in knowing what happens - my peculiar case ended up with best performance when the buffer size was only 256 bytes.

John Bowler <jbowler at acm.org>