On 14 Jan 2009, at 13:02, Burkhard Carstens wrote:
Am Mittwoch, 14. Januar 2009 04:50 schrieb Seth Grover:
I had the same problem. You could try to enable "BESTMATCH" in the
heap
manager by either compiling the rtl with "-dBESTMATCH" or changing
"{ define BESTMATCH}" to "{$define BESTMATCH}" in rtl/inc/heap.inc and
see if the situation improves. For me, it improved but didn't solve
the
problem completely. I had to create my own mem pool for some
frequently
allocated/freed structures.
In the general/average case, "best fit" produces slightly worse
results than "first fit" (which is the default). The reason is that's
you're more likely to end up with a bunch of unusably small blocks
over time than with first fit (there are of course usage patterns in
which this does not hold).
IIRC this was caused by usage pattern like this:
* free a huge chunk (a) of mem -> chunk is returned to heap manager
* allocate small chunk -> this results in heap manager splitting chunk
(a) to return the small piece (b)
* now allocating again a huge chunk (same size as (a)) results in heap
manager requesting a new chunk of mem from OS because the remainder of
(a)-(b) is not sufficient.
.. well, in short: memory fragmentation
With BESTMATCH enabled, the heap manager tries harder to find a small
free block for (b) before splitting (a) ..
However, I am not completely sure if this is the same problem ..
It probably is.
it
could also be caused by the reworked heap manager, which now handles
mem allocation per thread (if that's allready in 2.2.2 ??) ..
No, it is not. It will only be released in 2.4.0. And it will not
solve the problem, unless the differently-sized memory blocks are only
allocated in different threads.
You can also try using your platform's libc memory manager to see
whether it deals better with this usage pattern (add "uses cmem" to
the uses clause of your main program). Or, as mentioned above, use
your own memory pool (like the default memory manager already does for
small allocations).
Jonas
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
http://lists.freepascal.org/mailman/listinfo/fpc-pascal
