MapServer 3.6 slower than 3.5?

[dmorissette]

Hi Steve,

We've had a user report to us that MapServer 3.6 (release) was slower
than 3.5 (release).  We investigated this, and he was indeed right. 
MapServer 3.6 was approximately 50% slower than 3.5 on the same
mapfile.  We used gprof to find out that most of the time difference
comes from msImageFilledPolygon().

Our test was a set of 10 shapefiles (approx 10MB), and running it 15
times with gprof gave the following results:

                     Total time     msImageFilledPolygon()
  MapServer 3.5       34.2 sec          23.6 sec
  MapServer 3.6       47.3 sec          36.5 sec

The configuration options between the two versions were identical, same
copy of GD linked statically, etc.  Pointing to the same mapfile on the
same HD, doing a plain polygon fill in both cases.

Do you have any idea what may have changed to justify that difference? 
We can make the full set of gprof reports available to you if that can
help.

Daniel

[jlacroix]

The code of msImageFilledPolygon() and msPolygonLabelPoint() in mapprimitive.c
in mapserver 3.6 is now optimized. pointObj variable have been changed for
pointer of opintObj to avoid the duplication of variables and make the process
in 3.6 faster than 3.5.

Before the optimization, the function msImageFilledPolygon() made like half a
million of 32 bytes assigment with the line in a loop:
point1 = point2;

point1 and point2 were pointObj and now they are pointObj*

A couple of test is still needed to close the bug.

Note: this bug can be a duplicate of bug 154.

[dmorissette]

*** Bug 154 has been marked as a duplicate of this bug. ***

[dmorissette]

For the record, after Julien's changes to msImageFilledPolygon(), we went from 
MapServer 3.6 being 50% slower rendering polygon fills to it being almost 20% 
faster than 3.5!!!  This optimization has been applied to 3.6 and will be 
applied to 3.7 as well... further optimizations may be done down the road in the 
3.7 branch only.

Here are the new times for 15 iterations with the optimized code:

                     Total time     msImageFilledPolygon()
  MapServer 3.5       34.2 sec          23.6 sec
  MapServer 3.6       30 sec          approx 20 sec?


As Julien mentioned, there were two source code lines in critical areas of 
msImageFilledPolygon() that were doing copies of pointObj structures (i.e. 
point1 = point2;)... and that's where the 10 seconds of difference were spent.

In 3.5, pointObj was:
  typedef struct {double x, double y} pointObj;

In 3.6, we added measured shapefiles, and it became:
  typedef struct {double x, double y, double m} pointObj;

It turns out that this little increase in the size of the pointObj struct is 
what was costing us all that processing time.  After the code was changed to use 
pointers instead of copies of pointObj then we got the gain of performance 
described above (i.e. yes, 3.6 is really faster than 3.5 now!)

[dmorissette]

Oh... did I mention that 'gprof' is an amazing tool ?!?!?!?!?!?

[jlacroix]

The change in 3.7 are commited. I have finished the test and it does not appear
to have any problem. So, I'm closing the bug.