Re: Performance optimization (allocation inside a for loop)

[Jaroslav Hajek]

On Tue, Apr 7, 2009 at 3:48 PM, Francesco Potorti` <address@hidden> wrote:
>>I gave the idea a second thought - maybe it has some merits. Sometimes
>>you really want to use an array as a stack in Octave to avoid going
>>through a loop twice, so Octave can try to optimize such usage.
>
> Two thoughts only:
>
>  1) if stack operations are one target of this optimisation, then it
>    would better to have a hysteresis behaviour, that is, deallocate
>    one chunk only when the free space is 2*chunk big; more generally,
>    deallocate so that at least a whole unused chunk remains unused

Currently, the behaviour is that pop *never* reallocates. I think it
is not necessary, given the automatic stripping of over-allocated
portion when the array is being manipulated as whole.


>  2) 1kB for chunk size is very little today.  I would have considered
>    one memory page size at least, but since marginal improvements are
>    smaller for big chunks, maybe the current choice is good.  Only a
>    benchmark can tell, I suppose.
>

No problem. Feel free to do any benchmarks and suggest a better size
or even a better strategy. 1024 (i.e. 8kB for a double array) should
cut the amount of reallocations down by a factor 1000, so it seemed
enough to me.
There is still the possibility of making this chunk size configurable,
I only didn't do it because we already have hardwired help buffer
sizes elsewhere (blocked transpose, sorting).

regards

-- 
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz