Hyper-threading on or off for DCS?

[imacken]

Now, I know that there have been discussion on this before, but can someone please explain the attached images from Task Manager with DCS running?

The first is with HT off with a reasonable spread throughout the cores, and then the second is with HT on - my usual scenario. As I have observed many times, one core is almost permanently at 100% when running DCS.

As is the case with most tech topics, doing a search on Google gives very different opinions on whether HT is better off or on in games, but I just wondered what DCS users felt about this.

[imacken]

Can I ask why this thread gets moved to Bugs and Problems when there are no bugs or problems in the post, only an enquiring question?

[Harlikwin]

Well, I'm running a 9700k at this point so my answer is gonna be no hyperthreading since I don't have any. Sorry if thats not too helpful.

[imacken]

Well, I'm running a 9700k at this point so my answer is gonna be no hyperthreading since I don't have any. Sorry if thats not too helpful.

 

:lol: I've seen more useful posts from you, let's say that!

[Harlikwin]

:lol: I've seen more useful posts from you, let's say that!

 

I'm sure you've seen less useful too.

 

Looking at your runs, the non HT cores seem to working more. So I'd say that there is some efficiency gain by using HT for at least the one core case.

 

I assume this is the "same" test run in DCS. And does one case work better than the other?

Edited June 26, 2019 by Harlikwin

[BitMaster]

Indeed, do you see any differencies in flight ? FPS wise, stutter, other syptoms one has and the other doesnt ?

 

Tbh, the non-HT looks healthier to me as it has "some" headroom on the 1 core being stressed whereas the HT-setup has core #11 pretty maxed out. IIRC #11 is a HT core, #10 is the matching real core to #11.

 

FWIW, I use HT and have ZERO performance complains, really NONE.

Edited June 26, 2019 by BitMaster

[maxTRX]

Oh shoot! I forgot about HT. I recently flushed and updated my rig's bios and left the HT on. I'd rather have it off...:smartass:

[Rudel_chw]

For testing, I just disabled the Hyperthreading on my Rig and rerun this VR test: https://forums.eagle.ru/showpost.php?p=3953190&postcount=12

 

From an average of 42.3 fps I went up to 43.9 fps .. on same test conditions, so about a 4% extra increment .. not bad, so I will keep the HT disabled (I know, on AMD it has a different name, but dont want to confuse people)

[pimp]

Wow, I just turned off Hyperthreading in my bios and I got a real fps boost. I was around 70s-80s. I still have some ghosting issues even with ASW disabled. Also, the fps does seem like it gets stuck at 45, because if I look away and then back, the fps jumps back up to the 70s. I can do this every time unless it's around a congested area of the map. I don't have the time or patience to do the VR Test recommended by ED, but I have been flying Caucasus map with settings on high, MSAA 2x, & PD 1.5.

[Worrazen]

Now, I know that there have been discussion on this before, but can someone please explain the attached images from Task Manager with DCS running?

The first is with HT off with a reasonable spread throughout the cores, and then the second is with HT on - my usual scenario.

 

That spread through the cores in Task Manager's CPU view or similar does not necessairly indicate the parallelization of the software in question,, due to a normal default behavior of most thread scheduling algorithms which do not keep threads tied to a particular core, this isn't a feature or a purposeful thing, it is a phenomenon and without it you would probably get slowdowns in many situations/circumstances, I called this as"thread bouncing" initially when I was , but that's the whole point, they are bouncing all the time, but we who aren't the makers or really familiar with this don't know of the little fact that rarely only a certain core is responsible for a set for threads and that this bouncding happens at a very fast rate inside the CPU and because the CPU Core Utilization graphs work at a significantly slower refresh rate to show these changes in addition to the sample data being averaged out, this gives the appearance as if there are "two or more threads working on multiple cores"(work evenly spread out) while in reality it may just be be only one thread that was "thread bouncing" between two or more CPU cores.

 

So most of the time this kind of data presentation should not be used to gauge the parallelization of the software at all as it will be inaccurate most of the time. A different kind of data presentation would be needed which does not exist in the Task Manager so far, unfortunately there has been little to no effort from the industry such as MS and CPU manufacturers, probably because the industry actually benefits from the customers thinking as if everything they run on their new 8 core machines gets magically "parallelized", there is no such thing, I believe they are abusing this confusion in marketing ans PR when they say "oh that shows that your CPU is efficiently spreading the load", it is spreding it yes but if it's only one thread then that one thread isn't going to run any faster in practice (there is a separate debate whether it makes some bit of difference, it has to do with cache stuff, arhitecture type).

 

An application needs to built for parallelization (multi-threading) depending on the complexity and type of work even from scratch for the new multi-core CPU to be of any actual worth. Then it is a question how well is that implemented parallelized and also important is how parallelizable are the workloads that the application requires to reach it's goal. The per-core CPU Utilization graphs do show accurate information in practice, albeit it depends on their refresh/sample rate and how much rounding/averaging is used but if the person viewing it is not fully qualified in this area and knows exactly what he is looking at it will give a false impression, it will be wrongly percieved. Most of the end-users who are not industry insiders expect these CPU Utilization graphs that it will show them the activity also on a per-thread level, it does not show anytihng on a per-thread level at all. I wouldn't be surprised if even popular reviewers are fooled by this. Not saying it's necessairly MS's fault, but due to IMO massive confusion out there I think in their area that they should provide a warning to or some kind of disclaimer by adding a note directly into the Task Manager's CPU View, visible by default and and option to hide it maybe, to truncate it to an "i" info icon. One or two sentences wouldn't hurt, people need to be reminded about this. But this is only one fast and cheap remedy, the real solution is to actually provide another type of graph.

 

 

There is another behavior that is described as "core preferral", which means that some threads that are very active for most of the time have a tendency to stay on that particular core for that particular session or at least some amount of time. Whether or not this is only a scheduler thing or also an application software thing I do not know for sure, but this is what makes one core stand out of one CPU in the Task manager being at 100% ... as there's no need for thread bouncing that very active thread, rather other lesser threads bounce off that core so that one thread has all the space on that one core, that's the idea, althought that may not always be the case it could be a number of factors I have no idea about, scheduling is one complicated thing, and AFAIK it's proprietary, at least on Windows, well the code at least, I haven't gone into that area deeper since I got to the bottom of the local issue and I turned my focus back to DCS, but I might go and learn more about thread scheduling in depth in the future https://docs.microsoft.com/en-us/windows/desktop/ProcThread/scheduling-priorities

 

As I have observed many times, one core is almost permanently at 100% when running DCS.

 

This is normal, DCS isn't as parallelized yet, but it's more than I thought over a year ago, or it has changed since.

 

As is the case with most tech topics, doing a search on Google gives very different opinions on whether HT is better off or on in games, but I just wondered what DCS users felt about this.

 

When single-threaded operations are required for the work at hand, SMT (HT) will not generally not help. It will only help if there's many other smaller threads, for those to move away so the most important thread can have the most resources available on the physical core (can't say one particular core cause scheduler may bounce things ofcourse)

 

 

----

----

 

Also, with process lasso, I learned one thing just recently which should have been obvious, others may known it here much longer, there is a trick with CPU affinity about SMT (HT) that you can "disable" it per-process without having to do it in BIOS, all you have to do is disable half the cores, alternating, one after another is what I think the way to go (CORE 1 CORE 3 CORE 5 CORE 7, so that how you would prohibit the scheduler from putting the threads of a process onto multiple "hardware threads" (logical cores) on the same physical core.

Edited June 27, 2019 by Worrazen

[maxTRX]

Good info Worrazen… I disabled HT in bios after all this ruckus about new exploits that can be done in processors like i9 (I have one) and since it doesn't effect the performance negatively in DCS I'm keeping it off... for now. :/

[etherbattx]

HT means two threads can share a core. it only provides benefits in some workloads, but they can charge more for the feature and it’s a huge marketing point

 

16 threads sounds way better than 8 cores.

[JimmyWA]

Now, I know that there have been discussion on this before, but can someone please explain the attached images from Task Manager with DCS running?

The first is with HT off with a reasonable spread throughout the cores, and then the second is with HT on - my usual scenario. As I have observed many times, one core is almost permanently at 100% when running DCS.

As is the case with most tech topics, doing a search on Google gives very different opinions on whether HT is better off or on in games, but I just wondered what DCS users felt about this.

 

Hi Imacken,

 

You have to make sure DCS is running on more than one core.

 

Since the second last Beta update, on my PC it now randomly defaults to run on 1 (and I get 15 FPS in VR with stutters) or ALL (and I get up to 45 FPS with no stutters).

 

https://forums.eagle.ru/showthread.php?t=243170

[imacken]

I’ve seen your dozens of posts on this already!

I have never had that issue.

[BitMaster]

because he runs a WS chipset that suffered from this, X99, X299.

 

Z170/270/370/390 never suffered from this

[imacken]

Yep, we know that. He’s told us often enough now though!

[maxTRX]

OK, I'm one of those X299 chipset users and this HT stuff got me completely confused:huh:

In the initial bios config on my rig I tried to max everything at the same time (of course!) and I went the easy way by clicking 2 big "buttons" at the top of OC page: "game mode" and "XMS" for memory (3600 OC). The CPU was auto set to 4.1 and memory to 3600... fine. I ran like this for quite a while. Performance wasn't bad but could never get rid of stutters.

Recently I decided to disable HT that was initially set to on by "game mode". I went back in the sim with resource window open in task manager and saw all the cores active and DCS.exe had over 50 processes running on all these cores. How can I tell whether HT is off? Only physical cores are visible?

I opened bios again, disabled "game mode" and XMS. The CPU went back to default 3.3, confirmed HT was off but this time I also disabled Intel's "Turbo mode" as some posters here advised.

Back in the sim... what a shocker! No stutters whatsoever. FPS same as before (45 because of Fraps) but very smooth. After 20 min of flying (in maxed settings, PD at 1.5) there was one micro stutter. In Task manager I could still see between 50 and 60 processes running on all cores.

No more Turbo mode for me!

[Fisu_MAD]

WOOWW. HT Off Sure.

 

After being set at 40 FPS with my oculus S, now, depending on the height I can stay at 80 FPS when I look at the horizon.

 

With Hyper Threading on, it is not possible. They always stay at 40 FPS.

[JimmyWA]

I’ve seen your dozens of posts on this already!

I have never had that issue.

 

That’s good - because I asked ED to put it as a sticky but heard nothing, and I’m trying to make sure that anyone else suffering from this issue finds the solution.

 

I struggled for six months until I finally found this what the problem was, which has now been complicated by DCS randomly starting on 1 core. Hence me posting again.

 

A lot of people don’t hang around the forum but are sent to to one thread when they google.

 

Just because it isn’t a problem for you doesn’t mean it isn’t a huge problem for others.

 

I’m glad Ive gotten this information out there so well!

[imacken]

WOOWW. HT Off Sure.

 

After being set at 40 FPS with my oculus S, now, depending on the height I can stay at 80 FPS when I look at the horizon.

 

With Hyper Threading on, it is not possible. They always stay at 40 FPS.

 

Makes no difference with me. Performance is the same with or without HT.

[imacken]

That’s good - because I asked ED to put it as a sticky but heard nothing, and I’m trying to make sure that anyone else suffering from this issue finds the solution.

 

I struggled for six months until I finally found this what the problem was, which has now been complicated by DCS randomly starting on 1 core. Hence me posting again.

 

A lot of people don’t hang around the forum but are sent to to one thread when they google.

 

Just because it isn’t a problem for you doesn’t mean it isn’t a huge problem for others.

 

I’m glad Ive gotten this information out there so well!

 

I didn’t say it wasn’t a problem. Good that you got it out there, but your suggestion that is was ‘fixed’ in a DCS patch, I doubt very much. This has nothing to do with code in DCS.

My other gripe is that if we all posted the same point in every thread we visit, the forum would be out of control.

One thread should be sufficient for a Google search.

[toutenglisse]

I'm just posting a little monitored comparison between HT on and HT off. (Afterburner shows : %GPU, FPS, %CPU core by core - Edit : resolution 100ms) - this is in VR oculus CV1.

 

TF51D Free Flight Tibilissi - flying leveled and strait forward 30sec, looking forward.

 

Specs and DCS settings are the same - only HT is changed. ASW set to off

(Cpu 4.8Ghz without AVX clock offset - GPU stock clock (summer setting…) / + 405Mhz Vram)

 

[ATTACH]212969[/ATTACH] [ATTACH]212970[/ATTACH]

 

Result with these settings (most important being ASW off - no fps limiting) :

 

HT OFF - 4 cores - FPS are much more consistant / stable, always just above 70fps. GPU is fed perfectly - 100% all along the effective flight (inside the yellow Arrow).

 

HT ON - 8 cores - FPS are variating with low values during flight down to 59 fps (so less stable, almost equal to HT OFF at the second half of flight when city below starts to exit field of view). GPU is not fed as good as HT OFF flight.

Note : usually with HT ON I mostly see last logical core being maxed out and low usage on the rest of cores, but here with ASW OFF all cores are used.

 

Conclusion : HT OFF gives better performances for sure, but if you limit fps with ASW (or if ASW fail to be set to off like I used to experiment it before actual DCS OB) you maybe won't see a difference. (Edit : except in some heavy situations ?)

Edited July 1, 2019 by toutenglisse

[delevero]

Interesting scematics and conclution, thanks for sharing.

 

Note..: I see you are using the intel 7700 cpu.. Its a really good gaming cpu.

 

I do however like to add that people should be aware that HT with an

older model cpu compared to newer cpus that you might see other results since "modern/new" cpu's from example 2018 and forward uses HT and cores more intelligent than older cpu's do.

 

Example there is a big difference from the ryzen 7 1800x vs the ryzen 7 2700x since the 2700x uses its core in a more intelligent way.

Example it better slow down some of its 8 cores ( 16 threats ) and boost the speed of other cores in a better way.

( Currently pretty much all games on the market is developed so the most common gaming pc can run a game well, wich mean games are typical designed to run on 4 cores and the rest of the cores often are being used or not optimized since its really hard to program. So often do not take full advantage of all the cores ect.. Once vulcan engine get implemented it will be much better since it uses the hardware cores and gpus much better than directx does ) ;-)

 

Anyway

be aware that intel and different cpu's might not behave excactly as the intel 7700cpu in the test ;-)

Edited July 1, 2019 by delevero

[toutenglisse]

Yes, specs (hardware - age of gen.) and settings (of hardware and DCS) do impact results heavily.

[Yeti42]

Off, Off, Off or Off³