Re: I'm a total push-over..

On 1/23/08, Andreas Ericsson [off-list ref] wrote:

Dmitry Potapov wrote:

quoted

On Wed, Jan 23, 2008 at 09:32:54AM +0100, Andreas Ericsson wrote:

quoted

The FNV hash would be better (pasted below), but I doubt
anyone will ever care, and there will be larger differences
between architectures with this one than the lt_git hash (well,
a function's gotta have a name).

Actually, Bob Jenkins' lookup3 hash is twice faster in my tests
than FNV, and also it is much less likely to have any collision.

quoted

From http://burtleburtle.net/bob/hash/doobs.html

---
FNV Hash

I need to fill this in. Search the web for FNV hash. It's faster than my hash on Intel (because Intel has fast multiplication), but slower on most other platforms. Preliminary tests suggested it has decent distributions.

I suspect that this paragraph was about comparison with lookup2
(not lookup3) because lookup3 beat easily all the "simple" hashes
in my testing.  Only competitor was Hsieh one which was like 50:50
faster or slower depending on alignment / compiler / cpu.

---

My tests ran on Intel. I also noticed I had a few hashes commented out when
doing the test, one of them being Paul Hsie's. For some reason, Jenkin's and
Hsie's didn't perform well for me last time I used the comparison thing (I
did a more thorough job back then, with tests running for several minutes
per hash and table-size, so I commented out the poor candidates).

I still believe that for this very simple case, the lookup3.c case is not
very practical, as the code is that much more complicated, which was my
main point with posting the comparison. Iow, not "switch to this hash,
because it's better", but rather "the hash is not as bad as you think and
will probably work well for all practical purposes".

If you don't mind few percent speed penalty compared to Jenkings
own optimized version, you can use my simplified version:

  http://repo.or.cz/w/pgbouncer.git?a=blob;f=src/hash.c;h=5c9a73639ad098c296c0be562c34573189f3e083;hb=HEAD

It works always with "native" endianess, unlike Jenkins fixed-endian
hashlittle() / hashbig().  It may or may not matter if you plan
to write values on disk.

Speed-wise it may be 10-30% slower worst case (in my case sparc-classic
with unaligned data), but on x86, lucky gcc version and maybe
also memcpy() hack seen in system.h, it tends to be ~10% faster,
especially as it does always 4byte read in main loop.

-- 
marko