I did some profiling of flightgear because I'd like to optimize performance of the terrain culling + rendering which is bad when terrain fog is far, that mean when there is a lot to render and cull in the scene graph. Performance is fine with the default fog distance though but it gets very bad when you put it farther with keys Z/z. When fog is very far or disabled it can go below 20fps and become very jerky. Something I noticed also is that flightgear mem usage keep growing and growing over time and it is much more apparent and quick when loading a lot of tiles when fog is far. I had it to the point it filled quickly the swap and flightgear was killed by the kernel. I think most of the bad performance with lots of polygon implied by far fog comes from the scene graph from ssg. Another case of bad performance is terrain generated from 3 arcsec DEM with good resolution when the terrain is moutains or hills like Hell's Canyon (ftp://ftp.flightgear.org/pub/fgfs/Scenery/w120n40.tar.gz).
Test system:
Athlon 1900+ XP
512Mb RAM
Geforce 4 TI 4200
latest NVIDIA binary drivers 41.91
FlightGear compiled with:
- gcc 2.95.3
- FlightGear CVS
- SimGear CVS
- plib 1.6.0
- FlighGear and libs compiled with -g -pg
- no threaded tile loader support and no optimizations (-02)
Test run with options:
--disable-sound
--enable-auto-coordination
--aircraft=ufo
--start-date-lat=2002:01:10:11:00:00
--disable-clouds
800x600x32 windowed
ufo at max speed during 83 seconds running time.
all percentages are in % of total execution time
(part of the profile is in the attached file)
- 67.8% in fgRenderFrame() [2]
- 64.4% is from ssgCullAndDraw() [3] divided in:
- 12.9% : ssgCullAndDraw() called by SGSky::preDraw() that is
only used when skyblend is enabled (--enable-skyblend). It
is enabled by default
- 51.5% : ssgCullAndDraw() of the whole scene graph (terrain
+ objects)
or
- 53.3% culling
- 9.1 % drawing at the leaves of the scene graph
(ssgVtxTable::draw(void) [18])
- 20.7% in FGTileMgr::update() [8] most of which is due to
FGHitList::IntersectBranch(ssgBranch *, double (*)[3], double *,
double *) [13]
Some observations:
- most of the time is spent in ssg and sg functions
- I'm not sure the culling take 53.3% of total time because the
profile data is quite hard sometimes to understand because of cycles
but culling still takes a lot of time.
- drawing the blended sky shouldn't be taking 12.9% of total time i'll
see if I can fix that
- A fair amount of total time (14.1%) is passed in
sgFrustum::contains(sgSphere const *) const [15] the half of which
(7.05 %) is in the function itself and not its children (it might be
inlined function though).
prof.txt.gz
Description: GNU Zip compressed data
