Please sign in to comment.
Use a hybrid allocation scheme in xmlNodeSetContent
On Fri, May 11, 2012 at 9:10 AM, Daniel Veillard <firstname.lastname@example.org> wrote: > Hi Conrad, > > that's interesting ! I was initially afraid of a sudden explosion of > memory allocations for building a tree since by default buffers tend to > "waste" memory by using doubling allocations, but that's not the case. > xmllint --noout doc/libxml2-api.xml > when compiled with memory debug produce > > paphio:~/XML -> cat .memdump > MEMORY ALLOCATED : 0, MAX was 12756699 > > and without your patch 12755657, i.e. the increase is minimal. Heh, I thought that too. Actually you're looking at the result with XML_ALLOC_EXACT! This is because EXACT adds 10bytes "spare" on each alloc, and that interestingly wastes about the same amount of space as XML_ALLOC_DOUBLEIT on this example (see below). So it turns out that the default realloc() on my system actually handles this case really well — and I guess that all the time in xmlRealloc() was actually in xmlStrlen, not the underlying realloc() after all (sorry for misleading you). If you replace the realloc() with a bad one (like valgrind's), then the performance degrades severely. This patch implements a HYBRID allocator which has the behaviour you describe (it's like EXACT to start with, though without the spare 10 bytes; and switches to DOUBLEIT after 4kb) — that gets the memory back down to 12755657, with no noticeable impact on the performance of the synthetic pathological example under valgrind. In summary: max_memory on ./xmllint --noout doc/libxml2-api.xml, valgrind time on https://gist.github.com/2656940 max_memory valgrind time before | 12755657 | 29:18.2 EXACT | 12756699 | 2:58.6 <-- this is the state after the first patch. DOUBLEIT | 12756727 | 0:02.7 HYBRID | 12755754 | 0:02.7 <-- this is the state with both patches. > > There is also the cost of creating the buffers all the time. > I need to read the code and check but I may be interested in an hybrid > approach where we switch to buffer only when the text node starts to > become too big (4k would remove nearly all usuall types of "document" > usage, i.e. not blocks of data) I tried to avoid too much buffer creation by introducing the xmlBufferDetach function, which allows re-using one buffer to construct many strings. It's maybe a bit of a "hack" in API terms though I thought the gains would be worth it. Conrad ------8<------ To keep memory usage tight in normal conditions it's desirable to only allocate as much space as is needed. Unfortunately this can lead to problems when constructing a long string out of small chunks, because every chunk you add will need to resize the buffer. To fix this XML_ALLOC_HYBRID will switch (when the buffer is 4kb big) from using exact allocations to doubling buffer size every time it is full. This limits the number of buffer resizes to O(log n) (down from O(n)), and thus greatly increases the performance of constructing very large strings in this manner.
- Loading branch information...