Re: [xml] Useless function calls in xmlSetProp()?

[Julien Charbon <jch 4js com>]

[Argh! I joined a patch not a prog in my previous email. Sorry. This 
time I'm good. Just ignore my previous email]

Daniel Veillard wrote:
> On Fri, Feb 08, 2008 at 05:17:31PM +0100, Julien Charbon wrote:
> >  Seems fine and clear. Attached to this email the "final" patch 
> > against current trunk.
>
>   okidoc, rereviewed it, it looks fine, applied, tested still fine, so I
> commited it, thanks a lot !

  Thanks to you, for reviewing and applying patch.

>   In completely unscientific testings, runtest number of allocs are reduced
> from 3,058,476 to 3,053,663 which is around 0.15% , it really depends on
> the kind of documents used and what processing.

  Right. I made a tiny prog [again], that show time made by a call to
xmlSetProp() with an attribute value that double in size every iteration:

- With old xmlSetProp():

$ ./test-setprop-big
Size:   8       Time:   000:000014397
Size:   16      Time:   000:000003429
Size:   32      Time:   000:000003164
Size:   64      Time:   000:000004435
Size:   128     Time:   000:000033918
Size:   256     Time:   000:000017600
Size:   512     Time:   000:000042089
Size:   1024    Time:   000:000120426
Size:   2048    Time:   000:000400574
Size:   4096    Time:   000:001471683
Size:   8192    Time:   000:005275916
Size:   16384   Time:   000:020225833
Size:   32768   Time:   000:078549300
Size:   65536   Time:   000:304803839
Size:   131072  Time:   001:240064893
Size:   262144  Time:   004:963828528
Size:   524288  Time:   019:846718130
Size:   1048576 Time:   078:606054215

- With "new" [now current] xmlSetProp():

$ ./test-setprop-big
Size:   8       Time:   000:000004981
Size:   16      Time:   000:000001847
Size:   32      Time:   000:000000906
Size:   64      Time:   000:000000926
Size:   128     Time:   000:000001011
Size:   256     Time:   000:000001327
Size:   512     Time:   000:000001842
Size:   1024    Time:   000:000002458
Size:   2048    Time:   000:000004280
Size:   4096    Time:   000:000007559
Size:   8192    Time:   000:000022546
Size:   16384   Time:   000:000028533
Size:   32768   Time:   000:000054189
Size:   65536   Time:   000:000107761
Size:   131072  Time:   000:000218971
Size:   262144  Time:   000:000584088
Size:   524288  Time:   000:001057238
Size:   1048576 Time:   000:002148980

  [Yes, attributes with value size of 1 MB are unrealistic, it is just
to show how xmlSetProp() scaled before setprop.patch]

  Knowing these results, maybe there is a new possible performance
improvement:

  According on a quick run of valgring/callgrind on test-setprop-big,
old bad scaling is due to xmlStringGetNodeList() that call too much
xmlStrLen() through call sequence:
xmlNodeAddContent{Len}() -> xmlStrnCat() -> xmlStrLen().

  Thus, I will check if some improvements can be done
on xmlStringGetNodeList() to make it more scalable...

P.S: Joined test-setprop-big.c was compiled with:
$ gcc -Wall test-setprop-big.c -o test-setprop-big $(xml2-config
--cflags --libs) -lrt -g -O2

--
Julien

#include <stdio.h>
#include <stdlib.h>
#include <time.h>

#include <libxml/tree.h>
#include <libxml/xmlstring.h>

#define SIZE (1024*1024)
#define VALUE "value & "

struct timespec diff(struct timespec start, struct timespec end)
{
  struct timespec time;
  if ((end.tv_nsec-start.tv_nsec)<0) {
    time.tv_sec = end.tv_sec-start.tv_sec-1;
    time.tv_nsec = 1000000000+end.tv_nsec-start.tv_nsec;
  } else {
    time.tv_sec = end.tv_sec-start.tv_sec;
    time.tv_nsec = end.tv_nsec-start.tv_nsec;
  }
  return time;
}

int main(int argc, char *argv[]) {
  
  LIBXML_TEST_VERSION;
  
  xmlDocPtr doc = NULL;
  xmlNodePtr root_node = NULL;  
  
  struct timespec start, end, time;
  
  doc = xmlNewDoc(BAD_CAST "1.0");
  root_node = xmlNewNode(NULL, BAD_CAST "root");
  xmlDocSetRootElement(doc, root_node);

  xmlBufferPtr buffer = xmlBufferCreate();
  xmlBufferCat(buffer, (unsigned char*)VALUE);

  do {
    
    const xmlChar *value = xmlBufferContent(buffer);
    
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &start);  
    xmlSetProp(root_node, (unsigned char*)"attr", value);
    clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &end);
    
    time = diff(start,end);
    printf("Size:\t%d\tTime:\t%03ld:%09ld\n", xmlBufferLength(buffer), time.tv_sec, time.tv_nsec);
    
    xmlChar *tmp = xmlStrdup(value);
    xmlBufferCat(buffer, tmp);
    xmlFree(tmp);
    
  } while(xmlBufferLength(buffer) <= SIZE);
  
  xmlBufferFree(buffer);
  xmlFreeDoc(doc);
  
  return 0;
}