nvidia 256.52 xorg-server 1.9.0 performance regression - Page 3

Ralph · 09-07-10, 05:12 PM

What parameters are you passing to GtkPerf?

How many iterations?

bearoso · 09-07-10, 06:06 PM

Quote:

Originally Posted by Ralph

What parameters are you passing to GtkPerf?

How many iterations?

I think if you just run the program with the same configuration and parameters (or lack thereof) under both the older 1.8.x server and 1.9.0 server and compare them, it should give you a picture of whether you're getting the slowdown or not on your system. It doesn't really matter what other people are getting--differences in their choices of themes and software versions will skew the results.

On the other hand, anything drawing antialiased text with RENDER will be _noticeably_ slower if you're experiencing this regression.

torturedutopian · 09-08-10, 01:55 AM

>On the other hand, anything drawing antialiased text with RENDER will be _noticeably_
> slower if you're experiencing this regression.

Yep, it's definitely not necessary to use a benchmark to check whether this regression is occuring : drawing text gets *very* slow, which makes restoring windows, scrolling into windows, moving windows very jerky.

ioni · 09-08-10, 06:56 AM

i'm kinda disappointed about this issue.

We don't even know if is really an nvidia issue or an issue with xorg-server 1.9.0 because none of nvidia developers replied. My guesses are is an nvidia issue because with nouveau works

bearoso · 09-08-10, 09:30 AM

Quote:

Originally Posted by ioni

i'm kinda disappointed about this issue.

We don't even know if is really an nvidia issue or an issue with xorg-server 1.9.0 because none of nvidia developers replied. My guesses are is an nvidia issue because with nouveau works

It's not really anyone's fault, per se. The 1.9.0 server probably just hits a different path that isn't accelerated in the nvidia driver. Anything could cause this, but it's probably related to a memory format change. I even remember reading something a while back about using glyph atlases that might be related and could potentially be problematic to resolve.

The responsibility for fixing it is definitely nvidia's, though--it's up to them to find where it's falling back to software and hook the acceleration back up.

hl_ · 09-08-10, 10:01 AM

NVidia's driver could really need heuristics to deal with fallbacks in a more sane way. As it stands, if everything in the pipe can be accelerated it's REALLY fast, but if a fallback has to be used, it's REALLY slow, too slow. Other drivers seem to be able to decide better what should be done in certain fallback cases.

**AaronP** · 09-09-10, 04:17 PM

I tracked down the text performance regression to a change in the server's behavior, and have a fix ready that should be in a future driver release.

Nephyrin · 09-09-10, 04:36 PM

Quote:

Originally Posted by AaronP

I tracked down the text performance regression to a change in the server's behavior, and have a fix ready that should be in a future driver release.

Much appreciated!

torturedutopian · 09-10-10, 02:08 AM

Thank you Aaron, that was an important issue !

ioni · 09-10-10, 02:09 AM

Quote:

Originally Posted by AaronP

I tracked down the text performance regression to a change in the server's behavior, and have a fix ready that should be in a future driver release.

thanks a lot. beer for everyone!

Spyke · 09-12-10, 07:32 PM

I made the mistake of upgrading to Xorg 1.9.0 tonight with the latest Gentoo ~arch.

Even just using plain old fluxbox, everything with text draws horribly slow. You can literally see Xchat draw its chat box. Firefox pages scroll horribly as well.

I had to revert to Xorg 1.8.2 as its unbearably slow and unusable.

Hopefully that new driver is released soon. I expect it to be an annoyance for many when they too upgrade to Xorg 1.9.0.

transacid · 09-14-10, 03:34 PM

this is not really related, but could be. I'm running kernel 2.6.34 xorg 1.8.2 nvidia-drivers 256.53 and fluxbox. since some time (i can't really pinpoint it anymore) it takes longer for workspaces to appear when there is firefox running in it. It could be a firefox problem, but i wonder if anyone else experienced the same. gtkperf runs smoothly for me.

Code:

GtkEntry - time:  0,03
GtkComboBox - time:  1,74
GtkComboBoxEntry - time:  1,22
GtkSpinButton - time:  0,69
GtkProgressBar - time:  0,56
GtkToggleButton - time:  0,46
GtkCheckButton - time:  0,29
GtkRadioButton - time:  0,37
GtkTextView - Add text - time:  0,63
GtkTextView - Scroll - time:  0,34
GtkDrawingArea - Lines - time:  0,21
GtkDrawingArea - Circles - time:  0,34
GtkDrawingArea - Text - time:  0,22
GtkDrawingArea - Pixbufs - time:  0,14
 --- 
Total time:  7,23

09-07-10, 05:12 PM	#25
Ralph Registered User Join Date: Jul 2007 Posts: 74	Re: nvidia 256.52 xorg-server 1.9.0 performance regression What parameters are you passing to GtkPerf? How many iterations?

09-08-10, 01:55 AM	#27
torturedutopian Registered User Join Date: Oct 2007 Posts: 28	Re: nvidia 256.52 xorg-server 1.9.0 performance regression >On the other hand, anything drawing antialiased text with RENDER will be _noticeably_ > slower if you're experiencing this regression. Yep, it's definitely not necessary to use a benchmark to check whether this regression is occuring : drawing text gets very slow, which makes restoring windows, scrolling into windows, moving windows very jerky.

09-08-10, 06:56 AM	#28
ioni Registered User Join Date: Aug 2010 Posts: 63	Re: nvidia 256.52 xorg-server 1.9.0 performance regression i'm kinda disappointed about this issue. We don't even know if is really an nvidia issue or an issue with xorg-server 1.9.0 because none of nvidia developers replied. My guesses are is an nvidia issue because with nouveau works

09-08-10, 10:01 AM	#30
hl_ Registered User Join Date: Feb 2009 Posts: 138	Re: nvidia 256.52 xorg-server 1.9.0 performance regression NVidia's driver could really need heuristics to deal with fallbacks in a more sane way. As it stands, if everything in the pipe can be accelerated it's REALLY fast, but if a fallback has to be used, it's REALLY slow, too slow. Other drivers seem to be able to decide better what should be done in certain fallback cases.

09-09-10, 04:17 PM	#31
AaronP NVIDIA Corporation Join Date: Mar 2005 Posts: 2,487	Re: nvidia 256.52 xorg-server 1.9.0 performance regression I tracked down the text performance regression to a change in the server's behavior, and have a fix ready that should be in a future driver release.

09-10-10, 02:08 AM	#33
torturedutopian Registered User Join Date: Oct 2007 Posts: 28	Re: nvidia 256.52 xorg-server 1.9.0 performance regression Thank you Aaron, that was an important issue !

09-12-10, 07:32 PM	#35
Spyke Foxie Join Date: Aug 2004 Location: Canada Posts: 618	Re: nvidia 256.52 xorg-server 1.9.0 performance regression I made the mistake of upgrading to Xorg 1.9.0 tonight with the latest Gentoo ~arch. Even just using plain old fluxbox, everything with text draws horribly slow. You can literally see Xchat draw its chat box. Firefox pages scroll horribly as well. I had to revert to Xorg 1.8.2 as its unbearably slow and unusable. Hopefully that new driver is released soon. I expect it to be an annoyance for many when they too upgrade to Xorg 1.9.0. __________________ Gaming: Intel i7 980X @ 4GHz \| ASUS Rampage III Extreme \| GTX 480 3Way SLI @ 900Mhz \| Koolance VID-NX480 \| Corsair Obsidian 700D Corsair H70 CPU Cooler \| EK-FB RE3 \| Corsair AX1200 \| Black Ice SR1 360 \| 240GB OCZ Revodrive X2 SSD \| Windows 7 Ultimate 12GB Corsair Dominator GT @ 8-8-8-24-1T DDR3-1600 \| Onkyo TX SR-707 \| 70" Sharp Aquos LCD \| KEF Audio 5.1 C3/C6LCR/C7 Workstation: Intel i7 920 D0 @ 4GHz \| ASUS Rampage II Extreme \| GTX 480 @ 800Mhz \| Koolance VID-NX480 \| Lian-Li V1200B \| Corsair HX1000 EK NB ASUS HP \| Watercool HeatKiller 3.0 \| Feser Extreme X-360 \| 2x160GB Intel X25-M SSD RAID0 \| 4x2TB WD20EARS RAID10 12GB Corsair Dominator @ 8-8-8-24-1T DDR3-1600 \| Creative X-Fi Titanium \| Pioneer DVR-212D \| Gentoo Linux Dell 3008WFP \| JohnBlue JB3 \| CityPulse DA2.03e II DAC \| KingRex T20U w/ Modded Auricaps & PSU \| Glow Audio Sub One Server: Quad Socket Opteron 8356 (16 cores) @ 2.3GHz \| Supermicro H8QMi-2 \| Supermicro 2U Chassis \| Redundant 1200W PSU 32GB 2GBx16 DDR2-667 ECC \| 4x 300GB Cheetah SAS RAID10 \| Adaptec 5805 512MB w/BBU \| Colocated \| Gentoo Linux

09-14-10, 03:34 PM	#36
transacid Registered User Join Date: Oct 2008 Posts: 3	Re: nvidia 256.52 xorg-server 1.9.0 performance regression this is not really related, but could be. I'm running kernel 2.6.34 xorg 1.8.2 nvidia-drivers 256.53 and fluxbox. since some time (i can't really pinpoint it anymore) it takes longer for workspaces to appear when there is firefox running in it. It could be a firefox problem, but i wonder if anyone else experienced the same. gtkperf runs smoothly for me. Code: GtkEntry - time: 0,03 GtkComboBox - time: 1,74 GtkComboBoxEntry - time: 1,22 GtkSpinButton - time: 0,69 GtkProgressBar - time: 0,56 GtkToggleButton - time: 0,46 GtkCheckButton - time: 0,29 GtkRadioButton - time: 0,37 GtkTextView - Add text - time: 0,63 GtkTextView - Scroll - time: 0,34 GtkDrawingArea - Lines - time: 0,21 GtkDrawingArea - Circles - time: 0,34 GtkDrawingArea - Text - time: 0,22 GtkDrawingArea - Pixbufs - time: 0,14 --- Total time: 7,23