icefox tests issues

[Luciano Montanaro]

El Jueves 21 Octubre 2004 10:46, Lubos Lunak escribió:
> On Thursday 21 of October 2004 10:27, Juergen Pfennig wrote:
> > On Wednesday 20 October 2004 21:47, Simon Perreault wrote:
> > > Raul's results are very unsurprising and should be integrated with
> > > much rejoicing.
> >
> > Ok, I was always tinking of simple types. I am not a friend of operator
> > overloading in C++.
> >
> > But a factor of 5 (2 instead of 10s) just by removing some postfix
> > increments?
>
>  I agree. The results are very surprising, and most likely wrong. It
> would mean that just changing it++ to ++it in a loop removes 80% of
> executed code from the loop. It can make things faster, but such big
> improvement is hard to believe.

Well, it could happen if the operation to do on the container is simple 
compared to the overhead of the construction of the temporary iterator.
If these loops are called frequently - say, once per file scanned - the 
difference could be noticeable. So much difference is dubious however.
   
>
>  I suggest you repeat the test with exactly the same conditions for both
> of the variants (i.e. reboot, do the test; change to the other variant,
> reboot, do the test exactly the same way).

I think the test should be repeated a number of times in a row, so that the 
kernel buffers the filesystem reads, and the first iteraton should be 
discarded. Otherwise, the filesystem latencies would hide any difference 
between the tests.


-- 
Luciano Montanaro 

Any sufficiently advanced technology is indistinguishable from a yo-yo
                                                            - Enoch Root