On Wed, Jun 26, 2013 at 9:20 PM, Stephen Frost <sfrost@snowman.net> wrote:
> * Atri Sharma (atri.jiit@gmail.com) wrote:
>> My point is that I would like to help in the implementation, if possible. :)
>
> Feel free to go ahead and implement it.. I'm not sure when I'll have a
> chance to (probably not in the next week or two anyway). Unfortunately,
> the bigger issue here is really about testing the results and
> determining if it's actually faster/better with various data sets
> (including ones which have duplicates). I've got one test data set
> which has some interesting characteristics (for one thing, hashing the
> "large" side and then seq-scanning the "small" side is actually faster
> than going the other way, which is quite 'odd' imv for a hashing
> system): http://snowman.net/~sfrost/test_case2.sql
>
> You might also look at the other emails that I sent regarding this
> subject and NTUP_PER_BUCKET. Having someone confirm what I saw wrt
> changing that parameter would be nice and it would be a good comparison
> point against any kind of pre-filtering that we're doing.
>
> One thing that re-reading the bloom filter description reminded me of is
> that it's at least conceivable that we could take the existing hash
> functions for each data type and do double-hashing or perhaps seed the
> value to be hashed with additional data to produce an "independent" hash
> result to use. Again, a lot of things that need to be tested and
> measured to see if they improve overall performance.
Right, let me look.Although, I am pretty busy atm with ordered set
functions, so will get it done maybe last week of this month.
Another thing I believe in is that we should have multiple hashing
functions for bloom filters, which generate different probability
values so that the coverage is good.
Regards,
Atri
--
Regards,
Atri
l'apprenant
