At 05:00 PM 5/25/2003 +0400, Nickolay Samofatov wrote:

> > The screen shot is a little hard to read, but it's exactly what I would
> > expect. I
> > don't see BTR compress there at all. Did I miss something?
>
>1) Main idea is: scanning index bucket (BTR_find_leaf) is very long operation
>which usually consumes more then half of processor time during query
>execution. It would be good to change bucket structure a little to enable
>binary search and get rid of this overhead.

I won't say it can't be done, but prefix compression makes a binary search
virtually impossible since only the first node on the page contains the full
key. But if you have some ideas, let's discuss them.

>2) There is evidence that enabling SSE2 instruction set in compiler
>increases engine performance quite significantly even for Dialect 3
>databases without DOUBLE PRECISION columns. This probably means FPU is
>extensively used somewhere, but no profiling has been performed on
>this subject yet.

It's a very good idea to know what the problem is before offering a solution.
Get the evidence and we'll talk.

Jim Starkey