Re: Mixed 64-bit system GerH binaries / BillYscripts

On Sat, Aug 30, 2008 at 08:27:22PM -0400, Bill Yerazunis wrote:

> way back when  (where "way back when" was in the last millenium).  So,
> .css files used a pair of 32-bit numbers to simulate a 64-bit hash.
> That gained us about 5% accuracy improvement at the cost of doubling

but HS uses just 1 32b hash / bin and often outperforms the other 
classifiers

> But then OSB came up (thanks Christian!) and that cut the size of

That was Fidelis ;)

> Now a whole new millenium (and almost a whole new decade too) and
> now there's no need to use a pair of 32 bit ints any more,

besides being beyond 2k, do we have tests that clearly show the need or
a real overall improvement of hash+key/64b over just 32b, or conversely,
a significant real world (production) degradation ignoring the current
key field? 
I'd have expected hash clash effects to show off here, but in practical
use (real world, that is real msg stream) the 32b noise floor is well below
actual algorithm / method accuracy. 

> be necessary to make _all_ classifiers require two-sided training.

why?

> and all of these classifiers could be used together in the +same+
> CLASSIFY statement.  

ah, ok ... but then we'd need distinguished data chunks - a MIMO CLASSIFY?

> I'll post a high-level design and see what you folks think of it.

ok - but I favour from the start a compile-time selectable 32b/64b mode /
data-width, since 32b systems aren't going away anytime soon - likely no
more on desktops, but only on ultra-portables and PDA / *phones, and 
the latter is a huge 'market'.

-- 
paolo

-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/