Re: TOASTing smaller things

Hi Hannu,

On 3/22/07 3:21 PM, "Hannu Krosing" <hannu(at)skype(dot)net> wrote:

> Maybe monetdb ( http://monetdb.cwi.nl/ ) can be of some inspiration ?
>
> btw, it may be a good idea to have a verion of bizgresMPP which has
> monetdb as partition db, if monetdb is as efficient as they tell it is .

Yep - I've talked this over with the MonetDB folks in the past. The major
benefits they observe are those same things we talk about here,
implementation of long loops for operators and de-abstraction of operations
like compare() when appropriate, say comparing two INT columns in a sort.

We can get many of those benefits without by "vectorizing" the executor of
PostgreSQL even without the full column partitioning. We're in the midst of
working some of those changes as we speak. Early indications are that we
see large performance gains from this approach. Note that the actual
instruction counts per row don't change, but the more effective use of L2 I
and D cache and superscaler instruction units on the CPU create the big
gains. The MonetDB people present some effective literature on this, but
the important gains mostly come from the "vectorization", not the operator
de-abstraction IMO, which is good news for us all.

- Luke