Hyperthreading off works great!

[hqmurphy 7]

@dmiannay Thank you

[downscc 1,327]

I suggest that there be a pinned topic on HT and AM since it gets asked every other day.  Here is something I kept from earlier this year from the LM forum:

The default behavior is to use all cores. The only purpose of applying an affinity mask would be to disable one or more cores. We don't recommend this, but we recognize that it may be beneficial in some cases. For example, there could be an add-on that communicates with another application. There could also be other unrelated windows applications running, that users don't want to close when they run Prepar3D. Higher fidelity training devices often rely on additional applications to perform simulation work. In that case, the system might be configured to allocate specific cores to specific applications. This can be done via OS settings, but the AM setting in the cfg makes it a bit easier.

The other use case that is often mentioned is to disable the hyper-threaded core that shares a physical processor with the primary thread. Sometimes doing a full workload on the HT core can slow down the primary. This is really chip and workload specific. We updated our job scheduler in v2 or v3 such that it typically uses core 1 without overloading it. In our tests, even with HT on, using all cores was typically better than masking off core 1 after the scheduling changes were made. Since then, we have not recommended custom AM settings.

One last thing I should note is that add-on dlls are considered to be part of the Prepar3D process. The AM settings will be applied to those dlls as well. Add-on developers may be creating their own threads and assigning heavy work loads to them.

Beau Hollis P3D Software Architect
23 Apr 2018

End quote.

The interesting thing I noted is that in his reference to Core 1 (aka core 0) they have maximized the loading such that if you use HT then you have divided the workload assigned to core 0 into two logical cores.  My interpretation is that because full workload on an HT core (logical 0 and 1) can slow down the primary (logical 0) that HT on can depending on chip decrease performance.

 

Edited December 30, 2018 by downscc

[lownslo 592]

25 minutes ago, downscc said:

This is really chip and workload specific.

QFT!  There is no single solution.  For some running with HT off and no cpu masking works best, while others enjoy a better simming experience with HT on, an AM for the sim as well as all add-ons.  It takes time and testing to find what works best... some folks will take the time, and some folks just want to fly!  Whatever works for ya is what's best!

Greg

[SunDevil56 183]

18 minutes ago, lownslo said:

QFT!  There is no single solution.  For some running with HT off and no cpu masking works best, while others enjoy a better simming experience with HT on, an AM for the sim as well as all add-ons.  It takes time and testing to find what works best... some folks will take the time, and some folks just want to fly!  Whatever works for ya is what's best!

Greg

Pretty much...😎

Why some people just can't seem to come to grips with this fundamental reality has always been a mystery to me.

There's no one magic solution for everything and everyone in all situations, if there was we'd all be queuing up to buy it post haste.... lol

[mpw8679 428]

1 hour ago, downscc said:

I suggest that there be a pinned topic on HT and AM since it gets asked every other day.  

 

Good idea.  Just to add, include all the cfg. tweaks in the topic as well.

[Mace 1,909]

Yes P3Dv4.4 is behaving differently that 4.3 with my hardware.

With 4.4, in my limited testing I can indeed run with HT=off.  I could not do this in 4.3, as it resulted in a blurry mess.
Framelock at 30 (refresh=30), HT=off, I have no cpu overhead (core0 = 100% load) and sharp textures.
Framelock at 40 (refresh=30), HT=off, core0's load drops to about 75%-80%, essentially leaving me with some overhead, and sharp textures.

Unlimited framelock w/vsync and tb however, still gives me blurries with HT=off...even in a low load situation.

Regardless of where I framelock (but I must framelock, it appears), the results seem to be a much smoother experience in 4.4 with HT=off.  Lockheed Martin must have changed something.

My hardware is an 8700K overclocked to 5, 4K resolution at 30 mhz, 32 gb ram, 1080ti.  No AM tweak.

[IAhawkeyeDDS 158]

I’ve been tinkering with my 6700k overclock last couple weeks.

No difference in sim performance with HT on/off.  

 

4.6 GHz w/ HT on is semi stable (pass AIDA64, fail RealBench, pass p3Dv4.4 for 12 hours) but temps average high 70s with spikes around 85.

 

4.6 GHz w/ HToff is rock solid with average temps around 60 and no spikes into the 70s.

 

4.5 GHz w/ HT on is rock solid but I still see temp spikes into the 80s.

 

So I just roll at 4.6 w/out HT cause I like the temps.

 

just added 2080Ti.  It spikes into the mid 60s at times but it has really made a huge impact on alleviating stutters (upgraded from 1080 -non Ti).

 

my $0.02 anyways. 

[bean_sprout 1,201]

I think I heard somewhere that HT was going the way of the Dodo.  When all is said and done HT may have been a marketing thing.

Cheers
bs

[IUBrian 151]

I posted in another thread, my experience with HT off in my new 9900k/2080TI build made all the difference. I was going from an old 2500k/560 TI build and FSX and was giddy at moving to P3D. Fired it up and thought WHAT, same microstutters as with FSX and a system as old as Stonehenge, even at 5.0!? Turned off HT and the smoothness was the thing of dreams. Definitely not placebo effect. 

[Mace 1,909]

What is strange, is that in a low load situation (KJAC / Jackson Hole) my cpu usage was only 75-80% with HT=off, vsync on (30 refresh), yet a framelock of 40 and no AM.

Yet, with the same settings, at EGLL, I'm getting 100% cpu usage -- full load it seems.  Understandable given the number of scenery objects at EGLL, *but* why wasn't KJAC fully pegged on the cpu?  It was in 4.3 with no AM.

I experimented at EGLL by pulling cpu-based sliders back one notch but it was still pegged.  Maybe I'll keep pulling them back to see where usage drops below 100%, if ever.  This is strange because at KJAC cpu usage in 4.4 never hits 100% on Core0. I can only conclude that P3Dv4.4 does something differently with load balancing.

[glider1 90]

It is important to distinguish between P3D versions and CPU type. P3D 4.4 is different to 4.3 when it comes to core usage from what I'm seeing. I think it is also important to distinguish what operating system too, since the OS job scheduler is designating tasks to the least used cores and I think W10 does a better job at it than W7 did.

From testing I've done it appears core 0 is more concerned with allocating tasks to the other cores in 4.4 which means it is less busy but it gets more busy the more logical cores there are that need allocating. On a four core processor HT off, core zero is working roughly 10 percent less than with HT on in my tests.

Probably a six core processor is better with HT off now and definitely an 8 core. I'm not sure about a four core. I still am having better results with HT on with a four core processor but it is borderline unnecessary in 4.4 even on a four core processor it seems.

It is amazing to me that core usage has significantly improved in 4.4 something the community wanted for so long but it has gone through with almost no fan fare or hype and almost unnoticed by many for quite a while (including myself) and there is nothing in the change log either!

[Bert Pieke 7,584]

With each new release comes new learning...

V4.4. is pretty new, so you guys are all pioneers

[kevinfirth 1,393]

1 hour ago, glider1 said:

It is important to distinguish between P3D versions and CPU type. P3D 4.4 is different to 4.3 when it comes to core usage from what I'm seeing. I think it is also important to distinguish what operating system too, since the OS job scheduler is designating tasks to the least used cores and I think W10 does a better job at it than W7 did.

From testing I've done it appears core 0 is more concerned with allocating tasks to the other cores in 4.4 which means it is less busy but it gets more busy the more logical cores there are that need allocating. On a four core processor HT off, core zero is working roughly 10 percent less than with HT on in my tests.

Probably a six core processor is better with HT off now and definitely an 8 core. I'm not sure about a four core. I still am having better results with HT on with a four core processor but it is borderline unnecessary in 4.4 even on a four core processor it seems.

It is amazing to me that core usage has significantly improved in 4.4 something the community wanted for so long but it has gone through with almost no fan fare or hype and almost unnoticed by many for quite a while (including myself) and there is nothing in the change log either!

This.

Ditch your personal perceptions of how useful HT was - the basis on which they lay HAS changed because the way tasks are allocated across cores has changed as well.

10-20% increase in rendering performance by reallocating resources is great news - it means LMs software engineers REALLY understand how it works, AND that they're following a structured development path.

If you have six cores or more try experimenting with giving the second P3D task a physical core to itself as well as the primary render task.  

[Sartanius 399]

4 hours ago, kevinfirth said:

If you have six cores or more try experimenting with giving the second P3D task a physical core to itself as well as the primary render task.  

Right now, with an 8700K and HT Off, I have balanced loading across all 6 cores throughout a flight. No one core is more heavily loaded than another.   It is a thing of beauty to see. I don’t run any AM, just a stock configuration overclocked to 5.0GHz.

P3Dv4.4 working this way out-of-the box is what we’d all been hoping for, right? What am I missing here?

[kevinfirth 1,393]

Not all cores should necessarily be loaded in a balanced manner.

The primary task is the renderer, should aim for that to be 90-95% loaded, masking off the virtual core if using HT.

The second task does a lot of the managing workload, Steve recommended that be kept also to a physical core by itself.

The rest is scenery loading which will do work as and when necessary.

Anything that impact on the primary or secondary tasks will limit fps and/or induce long frames (aka stutters)