Ondemand governor dramatically slows down mesa perfomance

[brosis]

Using the EVE Online MMO as a test, here is a comparison between three environments. The OS is Debian 7.0 "Wheezy" on an AMD Phenom 9850 with a GTX260 and Nvidia 304.88 driver, kernel 3.9.3 and MESA 8.0.5.

I did not use the default WINE version that comes with Debian as it fails to run the game out-of-the-box with sound and used a locally compiled WINE 1.5.30.

Look, Nvidia is proprietary driver, okay?
Use opensource driver, for it is reproduceable and understandable of what parts in kernel do what.
We don't know if nvidia driver powermizer logic automatically forces CPU to clock high for such loads or not.

Your MESA is outdated, I also have no idea why are you even posting it, because its not used in the test. But lets ignore that.

WINE is using a LOT of CPU for DX to GLSL translations.
This also results in reduced performance (100 fps vs 300 fps).

Also, its an RPG, which means it uses a LOT of CPU by itself.

Multiply these two facts and try to understand by yourself why ondemand or performance play no role for this benchmark.. or why this game is very bad to test this specific issue...?

So,.. and also, ... based upon your 25 years of UNIX programming,... it would be nice you provide some code to fix the situation... or was that 25 years of UNIX trolling?

And finally, personal recommendation, sell that 260 GTX and buy something good. For example, I sold mine GTX 280 for 90$ and purchased HD5850 for 70$, which has near double the processing power at same TDP and which I now use with opensource driver and it works out great (aside from missing dynamic power management )

[brosis]

Obviously, as he fails to show any of the experience you credit him with. He fails to make a point, brings no evidence to his claim other than glxgears and lightsmark. I am looking forward to him stepping of the soap box and "hitting" me with some evidence. I am having coffee and pancakes by the way.

Okay, don't choke up. http://openbenchmarking.org/result/1...500792103#r-14

(windows) Games still run much better under Windows than under Linux.

No wonder, sherlock.

[sdack]

Look, Nvidia is proprietary driver, okay?

No, you look. I am not the one having a problem and I am also not other people's problem. Do not worry about me or my example. It is what it is - an example. Your problem is that you lack perspective. You do not see the bigger picture but instead continue to grasp for straws. Once you stop with it, have taken a deep breath and taken a step back might you get it.

I get that the ondemand governor sort of adds insult to injury to the entire situation around 3D gaming on Linux and that you want this to be gone. But you have to admit that it is not even up to the kernel developers to decide which governor to use for the default! You would need to talk to the distribution developers. No kernel developer will throw out the ondemand governor for you, or only accept a work-around, because of the decisions made by the distributions. What will happen is that the power management within the kernel will see newer features as part of the continuous kernel development and because of the advancements in technology. This may stop distributions from choosing the ondemand governor as their default, but it will remain their decision. It is then not a decision that was made by me and so your pleading for my understanding will only continue to fail. The kernel only provides a variety of tools, the distributions choose the most reasonable for their community of users, and the users decide which is best for their particular platform.

And about vadimg, if vadimg wants to add a work-around then he should do it. He does not need my consent, but only the consent of the other developers. So he probably should discuss it on the mailing lists. If he is looking for opinions then I have already given my opinion here and simply disagree.

Fair enough?

[sdack]

Okay, don't choke up. http://openbenchmarking.org/result/1...500792103#r-14

Thanks. I believe I have seen these before as part of a Phoronix article. Why are you showing me this?

Oh, I would only start choking if you made helpful comments without first suffering my comments.

[vadimg]

I doubt he is coming back. Once a game can run at speeds similar to Windows will the load average go up by itself and the ondemand governor will increase the core frequencies automatically.

Any single-threaded game won't magically start using all cpu cores, thus it will never result in 100% cpu utilization, and when the load is distributed among multiple cpus it won't always trigger highest frequencies. You don't even have to run a lot of tests or have 25 years of unix experience to understand that. Also rendering more frames per second probably means that gpu and kernel processing will also take more time, thus resulting in more waits on cpu side, so overall picture won't be very different from what we have today. So the problems are not going to disappear even with more optimized drivers, actually it's observable with the proprietary driver as well. Anyway, it doesn't really matter today whether the problem will disappear in the future or not, because people are affected by the problem *today*.

Obviously, as he fails to show any of the experience you credit him with. He fails to make a point, brings no evidence to his claim other than glxgears and lightsmark. I am looking forward to him stepping of the soap box and "hitting" me with some evidence. I am having coffee and pancakes by the way.

What evidence are you talking about? The evidence of performance issues observed with ondemand governor in some test configurations? Isn't it obvious from the very fact of existance of this thread and other comments reporting the issue here, including my own? If you think that the fact that you have no signs of this problem with some apps on your configuration proves that the problem doesn't exist at all, then your logic is completely flawed. You only proved that the problem doesn't exist in some test configurations, though nobody ever tried to claim the opposite in this thread.

And what 'his claim' are you talking about? I only provided some additional observations from my system and some thoughts about possible reasons, that's all. And then I mostly just tried to stop eruption of your advices.

He must have realised by now that I am right and know that there are a lot more actual bottlenecks inside Linux and that the ondemand governor needs no fixing. What do you think?

I never said that the ondemand governor needs any fixing, does it come from your own luxuriant imagination? It looks more like you are arguing with some imaginary opponent, because actually I didn't propose any changes/fixes to any component at all.

[peppercats]

Use opensource driver, for it is reproduceable and understandable of what parts in kernel do what.

I'd honestly be surprised if any 3D games actually launched with nouveau at a reasonable FPS.

[duby229]

hahahahahahaha. ha ha ha.

ok, this made me laugh.


From what i've seen, i doubt sdack will bother looking it up, so i'll just tell you: vadim is the primary developer who's been optimizing the r600g drivers.

Sorry if this spoiled everyone's fun. Just think of how many more insightful performance tips we could have gotten. No doubt sdack can let vadim know the best way of getting that extra 3% on all sorts of different games.

I guess you were right. Even telling him directly didn't seem to matter.

[sdack]

Any single-threaded game won't magically start using all cpu cores, thus it will never result in 100% cpu utilization, and when the load is distributed among multiple cpus it won't always trigger highest frequencies. You don't even have to run a lot of tests or have 25 years of unix experience to understand that. Also rendering more frames per second probably means that gpu and kernel processing will also take more time, thus resulting in more waits on cpu side, so overall picture won't be very different from what we have today. So the problems are not going to disappear even with more optimized drivers, actually it's observable with the proprietary driver as well. Anyway, it doesn't really matter today whether the problem will disappear in the future or not, because people are affected by the problem *today*.

What evidence are you talking about? The evidence of performance issues observed with ondemand governor in some test configurations? Isn't it obvious from the very fact of existance of this thread and other comments reporting the issue here, including my own? If you think that the fact that you have no signs of this problem with some apps on your configuration proves that the problem doesn't exist at all, then your logic is completely flawed. You only proved that the problem doesn't exist in some test configurations, though nobody ever tried to claim the opposite in this thread.

And what 'his claim' are you talking about? I only provided some additional observations from my system and some thoughts about possible reasons, that's all. And then I mostly just tried to stop eruption of your advices.

I never said that the ondemand governor needs any fixing, does it come from your own luxuriant imagination? It looks more like you are arguing with some imaginary opponent, because actually I didn't propose any changes/fixes to any component at all.

You are again picking out parts of my comments to show how little you understood them. To make matters worse are you derailing into pointless arguments. Let us see...

What has single-threading got to do with the ondemand governor in particular, and what has it got to do with the situation under Linux in general? ... Nothing! Each core can have their own governor. Nor do single-threaded applications present a problem particularly to Linux, but these applications are just as present under Windows and so make no difference to the discussion.
You then go on assuming - you are guessing and say "probably" - that higher frame rates will cause "probably" no change in the amount of CPU time spend in user space. If this was any true then changing a CPU's speed should also have no effect on the frame rate.
Instead do we have evidence of software, which produces 3 times higher frame rates under Windows, where it is not being throttled by the OS, than when run under Linux. So the CPU has to be executing game code three times more often or else it would not be producing the higher frame rate.
You then conclude that changes would not be very different from today.

Well, your statements are plain wrong. Games do run several times faster under Windows, and today, than they do under Linux and it is only because of the OS, and not because of the hardware or the games themselves.

You then claim that it does not matter whether the problem will disappear or not... Again wrong.

I will skip the rest. Just go back to your last comment where you replied with lots of little quotes and see what you wrote there.

Please read comments as a whole and do not respond before you have understood them. You cannot expect to produce meaningful responses without first having an understanding of what is being said.

[AnonymousCoward]

I'd honestly be surprised if any 3D games actually launched with nouveau at a reasonable FPS.

I had Quake Live running at close to maximum FPS (with some drops here and there) on a GTX275 a while back. Haven't had much luck with Nouveau working at all on my 660, though.

[vadimg]

What has single-threading got to do with the ondemand governor in particular, and what has it got to do with the situation under Linux in general?

I thought it's obvious, but I can explain: single-threaded application is the obvious example of application that doesn't fully utilize multicore cpu, so when the thread is moved between cores due to load balancing etc, utilization of any single core may be not enough for ondemand governor to raise its frequency, and driver optimizations won't be able to change this. In fact, most existing games (even not single-threaded) often do not fully utilize modern multicore cpus, so this argument is not limited to single-threaded applications, it was just a simple example.

Each core can have their own governor. Nor do single-threaded applications present a problem particularly to Linux, but these applications are just as present under Windows and so make no difference to the discussion.

There *is* a problem with many applications on linux, including single-threaded, otherwise this thread wouldn't exist. But ability to change governor per core and presence of single-threaded applications for windows has absolutely nothing to do with the problem that we see on linux.

You then go on assuming - you are guessing and say "probably" - that higher frame rates will cause "probably" no change in the amount of CPU time spend in user space. If this was any true then changing a CPU's speed should also have no effect on the frame rate.

I used the word 'probably' because nobody knows exactly how the behaviour of some random application regarding the waits, CPU/GPU synchronization, etc, will be changed with the *future* better optimized drivers and *future* kernel. Well, nobody except you.

But I didn't say that "higher frame rates will cause "probably" no change in the amount of CPU time spend in user space". Please quote my exact words (yes, that's why people use quotes) and answer to what I said, instead of attributing to me any phrase produced by your imagination and then answering to it.

By the way, didn't you notice that I said that the same problem is observable with proprietary driver? Of course you didn't, because this fact renders all your lengthy arguments regarding the future optimized open-source drivers meaningless.

Instead do we have evidence of software, which produces 3 times higher frame rates under Windows, where it is not being throttled by the OS, than when run under Linux. So the CPU has to be executing game code three times more often or else it would not be producing the higher frame rate.

Yes, CPU time spent in the game code is 3 times higher in this case. But you forgot that CPU also executes *different* driver code and this doesn't allow you to make any assumptions about the difference between the total (game+driver) cpu times. Ask yourself how higher framerate is achieved with better optimized drivers even with CPU-limited games where you can't get higher cpu utilization because it's already at maximum? Driver optimizations *reduce* CPU time spent in the driver per frame, thus *reducing* total (game+driver) cpu time per frame and allowing higher frame rate. So you get increased framerate but reduced cpu time per frame. Is the resulting cpu utilization higher with higher framerate? Maybe, but also it may be the same or even lower, especially if the driver optimizations convert CPU-limited application into the GPU-limited.

You then conclude that changes would not be very different from today.

Again, I didn't conclude that "changes would not be very different from today". Please answer to what I said, not to what you think I said. Why should I guess what 'changes' exactly do you mean when you attribute these words to me? I said "overall picture won't be very different", that is, there will be the same situation with underutilized cores, etc, and this all will result in bad performance with ondemand governor, same as today.

Well, your statements are plain wrong. Games do run several times faster under Windows, and today, than they do under Linux and it is only because of the OS, and not because of the hardware or the games themselves.

Did I really blame hardware or the games themselves for the difference in performance between linux and windows? I don't think so... Did I ever mention windows in my previous comment that you are answering to? Not really. So again, I'm not sure who are you talking with.

You then claim that it does not matter whether the problem will disappear or not... Again wrong.

What exactly is wrong here, could you clarify? Even if you are sure that that the problem will disappear in 3 or 5 years, how exactly does it help people with performance issues today?