Ragnar Kjørstad writes:
> On Wed, Jan 02, 2002 at 01:17:57PM -0800, Stephen Dennis wrote:
> > The approach is a combination of Alder32 and CRC-32. Alder32 is fast
> > for long requests because the compiler can parallize the internal
> > loop. However, it's relatively slow on short requests because of a
> > modulus operation that happens even on single-byte keys.
>
> Intuetively I would think that hash calculation time was almost
> neglectable compared to the other operations of opening a file (disk IO
> in particular). Has anyone done calculations or profiling to find out
> for sure if it's relevant?
"Tea" hash is ten times slower on some file creation patterns than
r5. This can hardly be explained by computational overhead, so, yes,
it's io caused by randomness that matters.
>
> > However, after exchanging E-mail with Jedi, it appears ReiserFS
> > needs similiar keys to produce similar but not exact hashes.
>
> Well, it depends on your accesspatterns - if you tend to create and
> access your files in a particular order there are (huge) performance
> improvements from a hash that sort the files in this particular order.
>
> If access is totally random, a well distributed hash will work best.
>
>
> --
> Ragnar Kjørstad
> Big Storage
>
Nikita.
