Variable simulation performance.

CPU: AMD 8320
GPU: NV 660

I noticed some pretty bad frame-rate drops around the towns in the east coast map, so I did a bit of investigation.

I drive around with a single car in the simulation, find the a point with the lowest fps, which in ~15. I check GPU load, which is ~25%, changed resolution, no difference in fps, so not GPU limited. I then disable shadows, reduce lighting quality, other graphical quality settings, no change either, so not graphics limited. I paused the simulation, and the fps increases to ~45, so obviously the simulation is extremely expensive.

So then I drive forward about 10m, and the fps goes way up, to ~35fps with the simulation running. I check the render stats, and there's no real difference between the two locations. I switch cameras, rotate around, no real change. I pause the simulation, and fps increases to ~45, just like in the other location. I drive back to the other position, sim off 45fps, sim on 15fps, just like before.

So with a single car, and without me changing anything, the simulation is vastly more expensive in one location than another, even though the locations are quite close to each other, and both are in the middle of a road.

Has anyone seen something like this? I can get more info if it would be helpful.

 

I noticed the same issue when I was creating The Nostalgia II track. I didn't figure out exactly what was causing it but it had something to do with the altitude of the road and bridges like you said.

 

It wasn't a bridge, but it was a road segment in a town that's built up over top of the regular terrain (i.e. if you turn on wireframe you can see the natural terrain way down below). Both locations were on a similar road, quite close together. FWIW I've actually noticed a similar framerate drop when crossing bridges.

I'm at work right now, but I'll get some screenshots later on.

 

Oh, here are some pics: http://imgur.com/a/itvkK

Each pair is with and without the simulation paused (J).

The first pair is in the town to the right of the spawn point on East Coast. With sim: 53ms, without sim: 19ms. So 34ms increase in the main loop just due to one car. Seems pretty bad.

The next pair is back down by the spawn point, looking out to sea. With sim: 17ms, without sim: 17ms. No impact on main loop. It's probably throttled or gpu limited here, but it proves that simulating this car should not take more than 17ms.

The final pair is on the bridge directly ahead of the spawn point. With sim: 48ms, without sim: 17ms. 31ms difference here.

I captured these before I saw your last message, otherwise I would have captured wireframes to go along with them. I'll go back in game and see if I can find any correlation with the collision mesh.

Cheers

- - - Updated - - -

I had a look at the collision meshes around where the performance gets really bad, and didn't notice any unusually large or small triangles, or any with extreme aspect ratios. The performance can be quite bad even when the car is sitting entirely in a single triangle. It's a bit hard to tell exactly due to the impact of the collision visualisation.

What is very obvious is the difference between driving on the static mesh and the 'ground model'. The performance always seems to be fine when I'm on the ground model (height field?), and there's always a 10-30ms hit when on the static mesh.

Let me know if there's anything else I can try to get you more info.

 

I tested again on 0.4.0.3 and everything seems about the same. Nothing in the patch notes indicated an improvement, but I thought I'd check anyway.