A Frame-Time Analysis of P3D v3 -- Effects of CPU Affinity, Frame Lock, and HT

[nooon 21]

AviatorMoser,

Great, great stuff. I would like to know what are your ASN settings during the bench tests. 

Could you dump your ASN Prepar3DOptions.CFG perhaps ?

[AviatorMoser 353]

Doesn't your scenery blur up at FFTF .01?  I tried this in FSX for fun and anything over 100 kts my scenery will blur and not ever load.  Of course I'm only at 4.0GHz it 2500K.

 

Not in the benchmark. I haven't tried anything else with FFTF tweak yet, such as airliner hop.

AviatorMoser,

Great, great stuff. I would like to know what are your ASN settings during the bench tests. 

Could you dump your ASN Prepar3DOptions.CFG perhaps ?

[General options]
FSVersion=Prepar3D
AutomaticDownloads=1
[Cloud options]
MaximumCloudLayers=5
MaximumCloudDrawDistance=130
[Audio options]
ApplicationSounds=0

[ryanbatc 12,820]

 

Not in the benchmark. I haven't tried anything else with FFTF tweak yet, such as airliner hop.

I forgot its just a small pattern.  I think with a low FFTF setting you can't practically use the sim.

[ADamiani 470]

..........I did have some hiccups over the western coast of Italy with Orbx OpenLC Europe. Not sure whats up with that. I tried to isolate the issue, disabled options, but never could find the source of stuttering in that case. ......

I confirm that! The western coast of Italy from Rome to Naples is the only place where I get stutters and occasional locks (around 1/2 a sec), otherways my 2600K + GTX 770 are quite happy with HT ON, AM=85, locked @30FPS and NI fixed at 1/2 refresh rate (60/2=30)

 

Rome is where I live, so it's a bit unfortunate ...

 

Andrea

[SAAB340 28]

Interesting text

Very interesting testing you've done there Daniel.

 

It's also interesting to see how very similar P3D is to FSX when it comes to certain things. Yes the engine seems to have been tweaked a lot so there are differences, but also similarities. I don't have P3D myself but really enjoyed seeing your more detailed analysis of its performance at different config settings being posted. I looked in to FSX performance in a similar way in the beginning of 2012. http://www.avsim.com/topic/377105-texture-loading-ssd-vs-hdd/ (The important images in that post are available on page two of the thread as the images in the original post got deleted).

 

The long spikes in frame times you are seeing appears to be of the same origin as in FSX. This is what I found out about them. "Adam gives a good clue in his terrain paper: "Since the rendering pipeline (Direct3D in this case) uses 32-bit floats, the final coordinates used for rendering must be converted to 32-bit offsets from a 64-bit local origin no more than several thousand meters away... ...As the viewpoint moves through the world, the local origin must be periodically updated to keep it within a few kilometers of the viewpoint. Otherwise, a loss of precision results making the terrain vertices appear to wiggle around as the viewpoint moves, an interesting but obviously undesirable outcome. Whenever we update the local origin, all the vertex buffers in every cell of the quad tree must be recomputed relative to the new origin. As this can take several frames, the new vertex data is double buffered and then made active only when all are ready to avoid visible cracks."

Its FSX terrain engine at work recalculating the vertex buffers. The 3 factors that have a big effect on this vertex update stutter, the most noticeable and longest of all stutters, are BP=0, LOD value and CPU speed. A faster CPU reduces this stutter."

 

The recalculation of the vertex buffers in FSX happens on the CPU as the GPUs of the time didn't have good FP64 support if they even had it. Given that the CPU speed still affects it a lot it seems like it's still done on the CPU in P3D. The higher the LOD, the more vertices you'll have and the more work for the CPU to keep them updated as you fly along. I believe BP=0 is already implemented in P3D.

 

A quick word about FFTF. Don't set it too low. I've looked in to effects of it in great detail in FSX and it looks like it behaves the same way in P3D. In a nutshell its a way to trade autogen/groundtexture loading against FPS.

 

A lower FFTF gives you higher FPS but lower autogen/groundtexture loading. Up to around FFTF=0.67 you get higer autogen/groundtexture loading at the expense of FPS. Settings above FFTF=0.67 don't yield any better autogen/groundtexture loading but still gives you even worse FPS.

 

If you however have FPS internally locked AND your CPU is at the moment sufficiently powerful to easily maintain the FPS lock you get maximum autogen/groundtexture loading even if the FFTF is set really low. That's why you don't always get blurries with a really low FFTF.

 

As the scene gets more demanding on your CPU a low FFTF let you keep your locked FPS longer before dipping (hence gives you higher FPS), at the expense of autogen/groundtexture loading.

 

In FSX you very rarely see a reduction in autogen, just blurry ground textures with a too low FFTF as you pretty much have to have a slow dual core CPU not to get full autogen. I suspect P3D might have a higher demand on more/stronger cpu threads as it has a lot more autogen.

 

It's interesting to see that locked frames inside P3D generates less microstutter vs unlocked. They have obviously worked on it. In FSX it was the opposite, but in FSX you can remove the visible microstuttering by using 1/2 rate v-sync and locking to 30FPS in fullscreen. (you can still measure it with FRAPS but what's actually displayed is a butter smooth experience as long as the vertex stutters don't get too high) I believe 1/2 rate v-sync does nothing for P3D as its fullscreen mode is actually a borderless Window and therefore 1/2 rate v-sync doesn't work.

 

Thanks for sharing your testing.

 

 

[jimcarrel 688]

Nice topic, I was using AM of 244, now using AM=85.

 

I7 3770 (non K) with frames locked and 30 and AM 244 I was running with a variation of .6 to .8%

with AM=85 it's dropped to a variation of .2% (running A2A' Comanche and ASN overcast skies)

 

This topic helped me squeeze out a wee bit more.

[Guest]

Tried the settings and a few more once again but in the end it's all the same... Perhaps a difference of 2 fps here and there but I still see microstutters when looking sideways and some major stutters every now and then. It's just how it is, I suppose. You can't compare P3D with modern games that run perfectly smooth. Unfortunately.

[ADamiani 470]

Tried the settings and a few more once again but in the end it's all the same... Perhaps a difference of 2 fps here and there but I still see microstutters when looking sideways and some major stutters every now and then. It's just how it is, I suppose. You can't compare P3D with modern games that run perfectly smooth. Unfortunately.

 

I'm afraid this post just takes the biscuit

 

Unless you factor in a serious overkill CPU+GPU pair, of course, or keep your settings REALLY down on a 30 Hz screen and all other recommendations that we already know.

 

All in all, I'm satisfied with what I have now, except, as I already said, somewhere around Rome, Naples and Florence in Italy.

 

But we all share the same feeling: that the improvements with V3 are substantial and the sim is working nicely now, but ......

 

Andrea

[DylanM 1,069]

Quick question to those still seeing stutters - are you using vsync and triple buffering? With my unlimited settings, I get no screen tearing and barely any long frames (unless I'm just approaching a heavy density area and the custom models are being loaded in...). I can share my setting if you like.

[AviatorMoser 353]

One thing I'm curious about is how you calculate the TEXTURE_BANDWIDTH_MULT value you use Daniel?

 

- JP

 

No calculation actually. I basically borrowed the value from what DoveTail used in the default FSX.cfg. I noticed FSX:SE had no delayed texture loading, so I just "borrowed" it.

[kevinfirth 1,392]

I also did the same with the copied TBM value - seems to work with no black textures

[Mickeyj 35]

I didn't have a chance to observe autogen issues. A five minute test probably is not enough time to ascertain whether autogen rendering is affected with the FFTF tweak.

 

Hello Daniel,

 

thanks for sharing your thoughts on AM frame lock and HT and of course the great Freemesh!!! I did som testing too in the last couple days messing around with AM and HT on/off. My CPU i7 5820k oc 4.6GHz + SLI 980 runs v3 best with HT OFF + AM=0 + FFTF=0.01.

I also tried some AM but experienced stutters. I also tried FFTF=0 this removed ALL autogen buildings only with consequently higher frames.

 

Check out my test i provided by request of Rob Ainscough http://www.avsim.com/topic/476776-500mb-vas-reduction-and-increase-fps-by-22-in-v3-need-some-testers/page-4 Her you find my results + settings + video footage.

[Penz 45]

No calculation actually. I basically borrowed the value from what DoveTail used in the default FSX.cfg. I noticed FSX:SE had no delayed texture loading, so I just "borrowed" it.

Thank you, I'm going to try that, my GPU is very similar to yours (GTX670).

[garrett_frank 150]

Is there a tl;dr version? I see a bunch of these types of, pardon the pun, threads pop up every now and then but there's never really a consensus. Which is odd because...

 

... it seems to me that the majority of simmers use an i7, such as the 2500k, 3770k, 4790k, 6700k, and so on. Given that they're all processors with 4 physical cores, should there be a basic rule of thumb at this point?

 

I ran some pretty scientific tests a while ago (not sure if I still have the results) but I found I was getting a 10% gain in FPS in a CPU-bound scenario with HT off in P3Dv2.5. I've never bothered with AMs.

 

So, what's the conclusion?

[Richard Sennett 2,499]

4790k - OC 4.6 - Titan X no OC - HT On

 

FPS Unlimited in P3D

 

This works great for me

 

NI = only this

AA mode = override any application

AA= 4X Sparse grid supersampling

 

 

[JOBSCHEDULER]

AffinityMask=84

 

[MAIN]

FIBER_FRAME_TIME_FRACTION=0.10