Hi Sergey

>Can you prove that and does it mater in practice?

I took Fabio's code and added a bit to make it easier to
check out in different scenarios. If you are using idx only as fast way to create temp sorts,
I found differences even greater than the 50% mentioned. Since the "typical" UC for
this probably is sorting grid columns, you just have to picture a user waiting <g>.
But this is dependant on the table size AFAIS.

For me the speed of lookup is more important (from my datamining stuff),
and here I found speed differences ranging from 10% to 30%.
Integer indices get the biggest boost -
probably a combination of selectivity and small key size.

I have to wait for a few other machines to run other stuff for testing across
a wider range of memory configs, but if you are using integer PK's
and have measurable a bottleneck there...

Haven't checked LAN perfomance, but the effects could be even greater there.

regards

thomas