Let me preface this by saying that I'm kinda perplexed by this result. I've messed around with affinity masks before in previous versions of Prepar3D and never had any conclusive results. Always some vague placebo "I think it's smoother if I do this..." but never the numbers to back it up. However, during my 'performance killer' run, I saw some major differences under v5. To explain, that run simply consists of taking off in a default jet (F-35 in this case) out of OrbX's SoCal and then fly for a few minutes. That will usually give you shaky FPS and stutters as the vast amounts of autogen cells are loaded. If the sim runs acceptable here, it'll most likely run acceptable everywhere else.

Under v5, I saw a huge difference. Turning HT off via an affinity mask, gave me a very smooth result, smoother than my sim has ever been here, whereas the default all cores mask was the usual shaky, low FPS like the previous entries (I'd say it was a good bit lower but my settings are also higher). Sim still stutters in both cases when a cell loads, which can be down to OrbX's models or just the sim of course.

I will also note, I had to use OBS instead of Nvidia Share, as that would make the sim crash with the RTSS overlay (It's a sensitive mistress this), so that could potentially have taken a bit of performance as well.

These are my settings and specs can be found in the signature. Framerate capped to 30 via Nvidia settings. 

 

Has anyone been playing around with affinity masks under v5 and seen any differences? 

Edited April 19, 20206 yr by Sethos

21 minutes ago, Sethos said:

Turning HT off via an affinity mask

Can you explain?  I am under the impression HT is controlled by a BIOS setting and in a then-booted system optimized within an individual application with Affinity Mask.

@Sethos I have an 9900KS. So 8 cores and HT on makes it 16. 
I added an affinity mask to disable the HT part of the cores. It runs way way way smoother. I tested it for 24 hours using the same scenario. And the conclusion is, that the hyper threading causes stuttering, and the affinity mask to turn it off removes the stuttering. 

6 minutes ago, fppilot said:

Can you explain?  I am under the impression HT is controlled by a BIOS setting and in a then-booted system optimized within an individual application with Affinity Mask.

Yes you are correct.  I disable HT in BIOS instead of setting it up as an affinity mask in the cfg.

8 minutes ago, fppilot said:

Can you explain?  I am under the impression HT is controlled by a BIOS setting and in a then-booted system optimized within an individual application with Affinity Mask.

It is but you can technically 'turn it off', which is what I'm referring to here, by simply masking off the logical cores, i.e only using the physical cores through an affinity masks. Then you end up with the same effect as disabling hyper-threading without having to do it globally.

I run HT on with an AM. Scenery loading (PS) vastly improved.

1 minute ago, Sethos said:

It is but you can technically 'turn it off', which is what I'm referring to here, by simply masking off the logical cores, i.e only using the physical cores through an affinity masks. Then you end up with the same effect as disabling hyper-threading without having to do it globally.

I would love to gain a better understand of that.  If you have time would you PM me a bit more about it?

Nothing new for my system. P3D v4.x was always a lot smoother with HT off. It's well documented that HT/SMT off can improve performance in some games as well, not all however. 

27 minutes ago, Sethos said:

It is but you can technically 'turn it off', which is what I'm referring to here, by simply masking off the logical cores, i.e only using the physical cores through an affinity masks. Then you end up with the same effect as disabling hyper-threading without having to do it globally.

That's not the same as disabling HT in the BIOS, because other running processes (there are hundreds) can still be tasked by the OS to run on the LPs that you masked-off in the P3D affinity mask. 

Consider this case:  HT on and an affinity mask that restricts P3D to run on only one LP per core (e.g. AM=85 or 01010101 on a 4-core CPU).  In this case, P3D's main thread will place a heavy load on LP 0, and no load on its paired LP 1.  The OS sees LP 1 as a good target to assign threads because it has no load.  So it puts threads from other processes on LP 1, which shares CPU's physical processing unit with LP 0.  You really don't want CPU 0 diverting its work from the main thread, but that's what will happen as the processing unit alternates between the instruction stacks on LP 0 and LP 1.  Additionally, with HT on, CPU 0 is sharing it's limited and super-fast cache memory between the processes running on the two LPs instead of using it all for the P3D main thread. 

With HT off in the BIOS, LP 0 is CPU 0...with it loaded way up by P3D's main thread, the OS will then assign other processes to the other less-heavily utilized CPUs.

 

2 minutes ago, w6kd said:

In this case, P3D's main thread will place a heavy load on LP 0, and no load on its paired LP 1.  The OS sees LP 1 as a good target to assign threads because it has no load.  So it puts threads from other processes on LP 1, which shares the processing unit on CPU 0 with LP 0.

Bob, that assumes that the Windows scheduler isn't aware of Hyperthreaded CPUs. It most certainly is, and has been since XP.

Cheers

5 minutes ago, Luke said:

Bob, that assumes that the Windows scheduler isn't aware of Hyperthreaded CPUs. It most certainly is, and has been since XP.

Cheers

And yet observation of LP 1 when running P3D on an HT-enabled CPU with an AM restricting P3D to a single LP per core does reveal that LP 1 is being tasked by the OS.

Same here, as for every version since FSX, best results on my rig are with an AM basically "turning off" HT. 

38 minutes ago, w6kd said:

And yet observation of LP 1 when running P3D on an HT-enabled CPU with an AM restricting P3D to a single LP per core does reveal that LP 1 is being tasked by the OS.

I think both can be true at the same time. 🙂

I'd be curious as to what processes were being scheduled and for how long. I'm not convinced that core contention or cache thrashing causes latency great enough to visible by the naked eye. I suspect manually changing the core affinity is either doing just enough to not trigger a specific race condition, or it's indirectly preventing said condition from happening.

Cheers!

What is noticeable in the video is the P3D5 Task Scheduler can easily bloat many cores on the left video at 100% for a certain amount of time, whereas on the right video, it seems to balance things out more evenly. It looks looks like it is non-HT aware and wrongly assigns tasks competing for the same resource on the same core which is mutually blocking both LP0 and LP1 at the same time.

I've never looked at whether P3D is using mainstream Microsoft CCRT, Intel TBB or their own task scheduler, but this video makes me wondering.

I noticed in the video, that with „HT off“ the GPU load is way higher than with „HT on“. So maybe P3D is shifting more load to the GPU when it has less CPU threads available, and the GPU handles this load actually better than the CPU?!