Aage Johansen wrote:

>
> Paul Reeves wrote:
> > Nico Callewaert wrote:
> > > And a page size of 4096, is that ok ?
> >
> > Not on the tests I did (very limited - networked client against an NT box w/
> > 64MB ram.) 4096 consistently came out worse than both lesser and greater values.
> > I do not understand why, but there is definitely a black hole there.
>
> 64MB isn't much for an NT box (if this is "TOTAL RAM" for the system).
> Do you have any measurements where the server had 128MB or more. What
> was the size of the cache (how many pages)? What was the amount of free
> RAM on the system before you connected to the database (and afterwards)?
>

64Mb isn't much. It was all I had to test with at the time. (Autumn 1999). I had
plans to do some reasonably scientific comaprisons of buffer size and
performance across NT, Linux and Solaris. Shortly after that the whole landscape
changed somewhat and the project died, along with a couple of others.

If you have the time and a suitable database I'd be happy to share what I was
doing with you. Suitable, in this case, is a database with two tables each large
enough to be bigger than the largest possible cache. This allows us to a) See
how long it takes to fill the cache and b) see the benefit of a full cache on
repeating the select. The second large table is used to flush the first table
from the cache.


Paul
--

Paul Reeves
http://www.ibphoenix.com
taking InterBase further