Performance- and Memory-Tests

1

I thought a bit about performance- and memory-(stress-) tests. And I plan to implement the following scenarios now:


  • very high number of fast rendering geoms (boxes and lines), without many changes to occur, flat tree  -> e.g. a graph browser
  • very high number of fast rendering geoms (boxes), with many changes to position and angle, flat tree  -> e.g. swarm simulation
  • high number of complex geoms but with few different meshes, flat tree -> e.g. asteroid field
  • very high number of fast rendering geoms in a deep tree (many nodes with a few children each), some of the nodes don't change, others update frequently -> (no example yet)



    These four have two phases - setup and runtime (separate mesurements). And they are the important ones, I think.

    Additionally they can all have view changes (on and off) and require with culling or not (objects near to each other or spread far apart).



    Others could be
  • loading many (different) meshes (maybe with partially same textures?)
  • flying over terrain, with and without loading/creating on demand
  • render to texture and altering textures programmatically?
  • other effects?


2

Eventually I'll make a test where objects are added and removed from the scenegraph frequently.

3

A stress test for RenderQueue would be nice. Perhaps a long line of objects that are randomly transparent and opaque. Have those objects move back and forth so the queue has to be sorted often.

4

I'm going to put them in jmetest.stress.[kind].

Where the test is named [Scenario]Test plus supplementary classes.

5
llama said:

Eventually I'll make a test where objects are added and removed from the scenegraph frequently.

I have written the 'swarm simulation' now and used reorganization of nodes to speed up collision detection - maybe this already suites as add/remove-test. I'll check it in as soon as I have cleaned it up.
6
irrisor said:

llama said:

Eventually I'll make a test where objects are added and removed from the scenegraph frequently.

I have written the 'swarm simulation' now and used reorganization of nodes to speed up collision detection - maybe this already suites as add/remove-test. I'll check it in as soon as I have cleaned it up.


Sounds perfect
7

a first test - swarm simulation - is in cvs

It is heavily using collision detection - perhaps we want to toggle behaviour on/off?

8

Certainly stresses the Mac  :-o



I get 0 FPS (maybe an update every 3-4 seconds).

9

Hmm, that's strange - it not fast over here but it's always over 10  (Pentium M 1,6 GHz)

10

Oh, uhm, not strange - I changed the Node class to speed up attach  :). is in cvs

11

ok, now getting about 3-4 FPS. At least it looks like it's doing something now. :slight_smile: If I hit R then it goes back to 0 FPS (maybe 1 update every 3-4 seconds).

12

TestGraphBrowser is in - about 4000 simple spatials getting rendered again and again.



Btw. this test could be speed up alot by automatically summarizing several meshes (lines in this case) into a single one…