Ondemand governor dramatically slows down mesa perfomance

[sdack]

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.

The ondemand governor measures the load within milliseconds and not seconds. If threads would migrate as fast and as often from CPU core to CPU core as you imagine it here then your primary problem would be the thread scheduler, because it is thrashing your caches. It is however not how schedulers work. You are only making up problems for not knowing any better.

All I am saying is that once you remove the actual bottlenecks will the ondemand governor see higher loads and fall in on its own. If these higher loads appear on one core or as an average over all cores is not the point.

You continue to misunderstand comments. I'll ignore the rest. Nobody wants to read this.

[vadimg]

The ondemand governor measures the load within milliseconds and not seconds. If threads would migrate as fast and as often from CPU core to CPU core as you imagine it here then your primary problem would be the thread scheduler, because it is thrashing your caches. It is however not how schedulers work. You are only making up problems for not knowing any better.

I'm not making up anything because this problem already exists. I never claimed that I know exact reason(s) though, and of course some of my initial ideas may be wrong (I don't even have 25 years of unix programming experience and I didn't seriously look into it yet), but it doesn't change the fact that there is issue that results in bad performance. That's why I just posted some ideas about possible reasons without making any conclusions or proposing any fixes, in hope that someone may have better explanation or any further ideas.

All I am saying is that once you remove the actual bottlenecks will the ondemand governor see higher loads and fall in on its own. If these higher loads appear on one core or as an average over all cores is not the point.

Sure, unless cpu usage will be still not high enough, or it will become even lower, but it's just a minor detail that of course doesn't affect significance of your conclusion.

You continue to misunderstand comments. I'll ignore the rest. Nobody wants to read this.

Looks like you already ignored all arguments that were against your opinion anyway, so ignoring anything else won't change much. If you really think that I don't understand your comments, you could as well simply avoid writing this comment at all and save some time.

[sdack]

I'm not making up anything because this problem already exists. I never claimed that I know exact reason(s) though, and of course some of my initial ideas may be wrong (I don't even have 25 years of unix programming experience and I didn't seriously look into it yet), but it doesn't change the fact that there is issue that results in bad performance. That's why I just posted some ideas about possible reasons without making any conclusions or proposing any fixes, in hope that someone may have better explanation or any further ideas.

Sure, unless cpu usage will be still not high enough, or it will become even lower, but it's just a minor detail that of course doesn't affect significance of your conclusion.

Looks like you already ignored all arguments that were against your opinion anyway, so ignoring anything else won't change much. If you really think that I don't understand your comments, you could as well simply avoid writing this comment at all and save some time.

I ignore most of your arguments, because they are a waste of time. Why are you wasting your time on non-issues when you could be working on the real bottlenecks? If you are a driver developer as your friends say then why do you not see improvements in the driver as the bigger win? We all can change the governor, but most of us are stuck with one and the same driver.

[ChrisXY]

If you are a driver developer as your friends say

This forum is great.

(Not a constructive comment, but then, what in here is anymore?)

[brosis]

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

G200a/b is modified G80, which is top-supported platform by nouveau; in other words it can be seen as a showcase of what nouveau is capable of.

@sdack
Welcome to my ignore list and go quack-quack somewhere else.

[smitty3268]

If you are a driver developer as your friends say then why do you not see improvements in the driver as the bigger win?

This really is the funniest thread I've read in some time here at phoronix.

[vadimg]

I ignore most of your arguments, because they are a waste of time. Why are you wasting your time on non-issues when you could be working on the real bottlenecks? If you are a driver developer as your friends say then why do you not see improvements in the driver as the bigger win? We all can change the governor, but most of us are stuck with one and the same driver.

Don't be so shy, I'm pretty sure that for such a great developer like you it would be easy to change the driver as well. In the worst case, if there are no more drivers to try, I have no doubts that you could write your own opengl driver in less than an hour and I'm sure it will be free of bottlenecks. And of course it will always result in 100% cpu usage to let the ondemand governor do its job.

[duby229]

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.


Under GNOME 3.4 is the frame rate 80 fps.


Switching from GNOME to Xfce 4.8 is the frame rate 91 fps.


Turning off all debugging inside WINE puts the frame rate at 98 fps. Notice that the frame rate itself becomes more stable.


Finally, switching to the performance governor makes no noticeable change other than my computer drawing more power from the wall.


The same game under Windows Vista. The frame rate here is about 370 fps. This only for a comparison.

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.

Awe, come on, the least you could a done was post a screenshot of something more exciting than a Noctis. Not knocking it though, it's the best salvager in game.... But, it's Eve...

[pheldens]

echo "15" > /sys/devices/system/cpu/cpufreq/ondemand/up_threshold # default 95
echo "10" > /sys/devices/system/cpu/cpufreq/ondemand/sampling_down_factor # default 1


Above works for me, it performs as good as the performance scaler with the advantage of powersavings and heat savings during single thread use on multicores. thanks for the tip everybody.

[RussianNeuroMancer]

http://www.phoronix.com/vr.php?view=19208