P3D v3.1 Performance Testing (RESULTS with 8-Core CPU) (Updated)

[Cruachan 705]

Hi Steve,

 

My understanding is that ProBalance Restraint temporarily lowers the priority of a process. It is difficult for me to understand why this should affect the consistency of the testing process if it only applies to IF10 Pro and not Prepar3D. Is it not true that following weather injection IF10 Pro's remaining function is to record the data generated by the TestFlight? If so, then why should this affect the validity of the results?

 

Have you managed to do any research into how Process Lasso functions? There is a useful FAQ which might help clarify any misunderstandings or misconceptions: https://bitsum.com/docs/pl/faq.htm

 

If it will help, I am prepared to uninstall Process Lasso temporarily and run a few more tests. However, I feel I would prefer to restrict these to those you consider to be essential to providing you with the evidence you seek as I found the whole process to be very time consuming.

 

It would also be helpful to see a video of your TestFlight, if that is possible. I know my results do not match yours and yet you will have seen that on 8 cores with HT disabled the on screen results are not too shabby at all.

 

Regards,

Mike

[SteveW 1,792]

I'm saying you can use the restore flight function in IF10 to restart and retest the exact scenario.

 

I'm also saying for the last time you cannot touch P3D or IF10 affinity, they do not work as intended.

[jabloomf1230 2,445]

Mike,

 

I've run P3d V3 like I stated in my last post (HT=off, AM only to isolate apps like ASN, etc.). I've used either Project Lasso or a batch file with Start commands and I've never seen any difference in either smoothness or frame rates between those two alternatives.

[SteveW 1,792]

 

The best thing to do to work this out is keep lasso off entirely. Start P3D from its desktop icon and attach with IF10, or better still simply start a saved flight with IF10 for repeatability and easy going. Set the affinity desired in the cfg. Pay attention to how addon exe's work, and be sure they are OK to lasso, many are not. When making graphs stick a dummy result on the end so all the graphs are made to the same scale. You're getting fairly good results with HT off and no AM, but big delta. You should be able to see better results, maybe down to 5% average delta. I've found results become more significant the more the sim is loaded, get big weather going, a mega airport and a PMDG for best results.

[Cruachan 705]

When making graphs stick a dummy result on the end so all the graphs are made to the same scale.

 

Hi Steve,

 

Sorry, I'm afraid you've lost me :(

 

Regards,

Mike

[SteveW 1,792]

Mike,

 

I've run P3d V3 like I stated in my last post (HT=off, AM only to isolate apps like ASN, etc.). I've used either Project Lasso or a batch file with Start commands and I've never seen any difference in either smoothness or frame rates between those two alternatives.

Jay, Mike's using specially written software that can start and monitor P3D and also respect its affinity like no other app. It measures the change in fps which sawtooths when there's a problem. In all of Mikes graphs the delta sawtooths, so that's a poor result. In your tests, there could be any number of things capping your fps, but it need not be, because anyway, no AM + HT disabled is always a consistent setup. It's just not the best use of the chip, logic should dictate that. You've got HT disabled? C'mon guys, you've got to work harder than that...

 

 

Sorry, I'm afraid you've lost me

Well, it's not for trying. Why are you even bothering introducing another app into the fray? You are trying to extract results from P3D make it work hard in as simple a way as possible. P3D does not work within the walls of a lasso process - period.

[SteveW 1,792]

Mike, Your best results shown are when the affinity mask uncovers the entire CPU and HT is disabled. Two simple reasons; since the affinity isn't narrowed it does not aggregate threads onto less cores, and HT off simply means there's half the threads per core in the first place. Doubling threads on a core is obviously going to task it more, and turning off HT halves that problem, but that's all it does. With P3D Rob pointed it out in other threads, there's processes going on outside the affinity that are supposedly attributed to the sim. If you narrow the affinity, these get bunched onto each other's cores and increase load on those cores.

 

If you lasso affinity of P3D, or IF10, then by definition it is constrained within an affinity space you defined for it, irrespective of how it thinks it starts. That process does not reflect the working relationship the sim is designed to expect when run on these PCs. The real problem for these apps (FSX/P3D and even XPlane) is enabling HT, they are simply fooled into creating twice as much work per core as they do without HT.

[Cruachan 705]

Hi Steve,

 

1. With respect, can I ask you to read post #20 again. You didn't really respond in the way I had expected. I just needed some clarification to "When making graphs stick a dummy result on the end so all the graphs are made to the same scale".

 

2. What is ExpandedFSAffinity (EFSA) and why is the default IF10 Pro setting=True?

 

Regards,

Mike

[SteveW 1,792]

Say for example you are getting results around 50-70 on y scale, if you always add a 90 on the end of the data, Excel will make the graphs with the same scale each time. EFSA will make your results worse if affinity is touched, except if you run with no AM HT disabled.

 

In this example you can see the constant dummy value of 60 needed to be 70 so all the graphs y scale come out the same. Also note that you should be seeing similar results where HT enabled is marginally better than HT disabled as in these results:

 

[Cruachan 705]

Ah, I see. So that explains the spike at the end of the 'Average fps' graph line in your images. I did wonder what caused it.

 

Mike

[SteveW 1,792]

If you are not seeing HT=on performing better, there's a problem with your setup. If you're seeing sawtoothing of delta >5fps there's a problem. My top graph shows an average around 2-3fps.

 

When you got things right, EFSA will help matters when things change during the game, for example if you connect with an external simconnect app (initiates a new client handler within the sim).

 

...Make sure Excel is not open on the log file .csv, since Excel holds the file open exclusively and results cannot be added to the file in that case.

[SteveW 1,792]

I've not tested in isolation, results came in the same with twin xeons. I concur with Intel, an HT LP always beats an non-HT LP since there's always more on that core than just one single thread. Logic alone suggests that these results (as I've shown above) are going to look almost the same (as they do). But your results show marked differences Mike, and show HT enabled as a fierce problem.

 

The background threads of P3D are broken into fibres. These task switch on the LP of that job, so they don't get any benefit from HT enabled. But they don't lose out to it unless they are also spawned on the companion HT LP of that core - forcing unnecessary sharing of that core.

 

What you should be seeing is that as the sim is loaded up more, the four core setups (HT on or off) should pull away from the five core setups more obviously.

 

There's been plenty of posts around the site whereby someone goes into task manager, set affinity, and need only press OK, don't even change anything, and they get more activity. It is this extra activity on the cores that increases the delta.

[Cruachan 705]

Hi Steve,

 

Can I say, once again, how much I appreciate the time you are spending on this topic. Clearly there is much more going on than we have perhaps been assuming.

 

Your graphs do not include Frame Rate (fps) and Delta fps. Can you explain why other than to declutter? It would be interesting to see how your frame rates fluctuate during testing as compared to mine. With Hyperthreading=On or Off with a 4-Core Affinity Mask mine could be likened to a 'perfect storm'! Your graphs are showing averages and not the raw 'behind the scenes' activity.

 

Twin Xeons, eh! I wonder whether Jay could be persuaded to run a couple of tests and post his results?

 

Regards,

Mike

[SteveW 1,792]

I find the averaged items more easily show the trend between settings at the end of those two minutes, trends which continue the same after four minutes. One minute tests are useless since there's concurrent tasks going on for over a minute that need to settle down after the sim starts. Maybe you could make a few three or four minute tests to check how it's going. Other than that I'd basically decluttered in that example.

 

One thing's for sure, given our polar opposite results, the program takes readings and seems to show no favouritism for any particular setup.

 

Remember the maths cannot lie. If one delta trace shows a better average than another, that is because the change in fps is less in that trace.

[Guest]

Steve/Mike,

 

This is with HT OFF and for this test I ran a more common 4.3 GHz (not my usual 4.6GHz).  Graphics settings are in the video (at end) and I have over 150 add-ons installed.  The airport is FSDG's Thessaloniki with surround area, rest is Orbx, REX, etc. etc. 

 

 

There are a couple of long frames here and there (but graphics settings are pretty extreme and I used the most complex time of day), but so long as one can achieve their monitor's HZ or their target FPS during most stressed rendering, then long frame frequency should be minimal (masked).

 

I've ran this same test with many permutations of HT ON/OFF AM's etc. ... even disabled cores ... and I agree with Steve's delta findings.  But 8 cores HT OFF no AM seems to be the best option for me and my 5960X at 4.3Ghz and/or 4.6Ghz.

 

I understand that keeping the deltas less is important to smooth, but what needs to be considered (by end user):

 

1.  Lower delta but overall lower FPS that is more consistent

2.  Higher delta but overall higher FPS with occasional (or in some cases frequent) long frames

 

So if my system is running at 24 fps with no long frames, or at 30 fps with occasional long frames (stutters), or 35 fps with many long frames (stutters) which is ultimately "better".  End user decides what is "better", some users may prefer the 24 fps consistent approach (low delta), some users may prefer the 30 fps with occasional stutter.  The end user needs to balance that out what's important to them.

 

When I disable vSync and tripple buffer I do get much higher deltas with HT OFF ... but that's not necessarily a "bad thing", it does mean the CPU is getting stuff done faster overall ... with vSynch and triple buffer that "masks" the deltas, but I'm ok with that because my CPUs are working as fast as they can.  The problems start to surface when one drops below one's target FPS and/or monitors refresh rate ... that's when one really needs to evaluate the importance of the delta of which Steve has done a great job demonstrating.

 

I know we've had many of these threads, but I think they are good threads and I know people have used Steve's work (I've pointed many folks to Steve's input) to help them achieve their goals.

 

Cheers, Rob.