Re: 1.4 - 1.5.1 performance regressions
On Tue, 2009-01-27 at 12:42 +0100, Fabio wrote: On Tue, 2009-01-27 at 11:55 +0100, Fabio wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I noticed that pre-1.6 is faster than 1.5.3, but still slower than 1.4. As reported by someone in the bug report, the patch from https://bugs.freedesktop.org/show_bug.cgi?id=16647#c11 was never applied. Is it still needed? Have you tried it to see if it helps for the performance regressions you noticed? No (I am using prebuilt packages), however, according to this comment, it fixes the regression: https://bugs.freedesktop.org/show_bug.cgi?id=16647#c14 ... for the benchmark / workloads discussed in the bug report. Are the workloads you noticed the performance regression with similar to those? I am getting these results with gtkperf, compared to what I got with 1.4. Note, however, that not all performance differences could be due to the X server, since I upgraded all system during that time (the radeon driver however is almost the same): + - faster - - slower = - about the same 1.4 1.5.99.901 Difference GtkEntry - time: 0,090,09= GtkComboBox - time: 3,835,74- GtkComboBoxEntry - time: 2,964,53- GtkSpinButton- time: 0,510,53= GtkProgressBar - time: 0,270,24= GtkToggleButton - time: 1,340,50+ GtkCheckButton - time: 1,440,28+ GtkRadioButton - time: 1,790,77+ GtkTextView - Add text - time: 11,68 11,96= GtkTextView - Scroll - time: 2,842,22+ GtkDrawingArea - Lines - time: 2,434,90- GtkDrawingArea - Circles - time: 1,863,77- GtkDrawingArea - Text- time: 3,893,42+ GtkDrawingArea - Pixbufs - time: 0,300,36- --- Total- time: 35,22 39,33- See also https://bugs.freedesktop.org/show_bug.cgi?id=16647#c24 ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Wed, 2009-01-28 at 16:20 +0100, Fabio wrote: I am getting these results with gtkperf, compared to what I got with 1.4. Note, however, that not all performance differences could be due to the X server, since I upgraded all system during that time (the radeon driver however is almost the same): If the GTK version / theme is different, that could indeed easily make a difference. Without trying the patch, it's hard to say what its impact will be on your numbers. See also https://bugs.freedesktop.org/show_bug.cgi?id=16647#c24 dixLookupPrivate should now essentially be an array lookup again, so any overhead this still incurs should have been there in 1.4 already, unless EXA calls dixLookupPrivate significantly more often now... I'll look into further reducing this by passing around pointers to the EXA private data directly internally, but I'm not sure how far that will go. And even if this overhead was removed completely, 10% speedup would be barely noticeable in practice if at all. -- Earthling Michel Dänzer |http://www.vmware.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Tue, 2009-01-27 at 11:55 +0100, Fabio wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I noticed that pre-1.6 is faster than 1.5.3, but still slower than 1.4. As reported by someone in the bug report, the patch from https://bugs.freedesktop.org/show_bug.cgi?id=16647#c11 was never applied. Is it still needed? Have you tried it to see if it helps for the performance regressions you noticed? FWIW, Something like it would probably still be useful to prevent the pixmap valid regions from growing too many rects. I've been meaning to clean up and push it, but haven't got around to it; hopefully soon. It's probably too late for 1.6 anyway though. -- Earthling Michel Dänzer |http://www.vmware.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Tue, 2009-01-27 at 11:55 +0100, Fabio wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I noticed that pre-1.6 is faster than 1.5.3, but still slower than 1.4. As reported by someone in the bug report, the patch from https://bugs.freedesktop.org/show_bug.cgi?id=16647#c11 was never applied. Is it still needed? Have you tried it to see if it helps for the performance regressions you noticed? No (I am using prebuilt packages), however, according to this comment, it fixes the regression: https://bugs.freedesktop.org/show_bug.cgi?id=16647#c14 ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Tue, 2009-01-27 at 12:42 +0100, Fabio wrote: On Tue, 2009-01-27 at 11:55 +0100, Fabio wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I noticed that pre-1.6 is faster than 1.5.3, but still slower than 1.4. As reported by someone in the bug report, the patch from https://bugs.freedesktop.org/show_bug.cgi?id=16647#c11 was never applied. Is it still needed? Have you tried it to see if it helps for the performance regressions you noticed? No (I am using prebuilt packages), however, according to this comment, it fixes the regression: https://bugs.freedesktop.org/show_bug.cgi?id=16647#c14 ... for the benchmark / workloads discussed in the bug report. Are the workloads you noticed the performance regression with similar to those? -- Earthling Michel Dänzer |http://www.vmware.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. -- Earthling Michel Dänzer | http://tungstengraphics.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I can't update at the moment (I am using prebuilt packages from Ubuntu). Can you point the fix? I don't see it at: http://cgit.freedesktop.org/xorg/xserver/log/?h=server-1.5-branch ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
1.4 - 1.5.1 performance regressions
Hi, I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. As a benchmark I am using GtkPerf [0] with gtkperf -a -c500 without compiz on a MacBook Pro with a Radeon RV530. While I noticed some improvement going from 1.5 to 1.5.1, 1.5.1 is still slower than 1.4. This may be related to recent report of slowness (thread http://lists.freedesktop.org/archives/xorg/2008-October/039164.html). My reference systems are: 1) 1.4 system: installed system with Ubuntu 8.04 with xserver 1.4, kernel 2.6.24 + backported -ati driver, radeon drm module, libdrm, mesa (used many recent git version); 2) 1.5.1 system: LiveCD Ubuntu 8.10 with xserver 1.5.1, kernel 2.6.27-rc9 recent -ati git snapshot, radeon drm module from kernel, libdrm 2.3.1, mesa 7.1. Does someone is able to reproduce this regressions, maybe changing only the xserver? Is there some known bug? [0] http://gtkperf.sourceforge.net/index.php?page=download GtkPerf benchmark results: 1.4 system, XAA: GtkEntry - time: 0,09 GtkComboBox - time: 4,10 GtkComboBoxEntry - time: 3,47 GtkSpinButton - time: 0,50 GtkProgressBar - time: 0,27 GtkToggleButton - time: 1,48 GtkCheckButton - time: 1,55 GtkRadioButton - time: 1,88 GtkTextView - Add text - time: 11,63 GtkTextView - Scroll - time: 7,23 GtkDrawingArea - Lines - time: 1,13 GtkDrawingArea - Circles - time: 1,84 GtkDrawingArea - Text - time: 29,33 GtkDrawingArea - Pixbufs - time: 3,57 --- Total time: 68,07 1.5.1 system, XAA: GtkEntry - time: 0,15 GtkComboBox - time: 6,81 GtkComboBoxEntry - time: 3,57 GtkSpinButton - time: 0,99 GtkProgressBar - time: 0,49 GtkToggleButton - time: 0,83 GtkCheckButton - time: 0,44 GtkRadioButton - time: 1,47 GtkTextView - Add text - time: 13,45 GtkTextView - Scroll - time: 2,52 GtkDrawingArea - Lines - time: 2,49 GtkDrawingArea - Circles - time: 3,39 GtkDrawingArea - Text - time: 45,16 GtkDrawingArea - Pixbufs - time: 4,36 --- Total time: 86,12 1.4 system, EXA: GtkEntry - time: 0,09 GtkComboBox - time: 3,83 GtkComboBoxEntry - time: 2,96 GtkSpinButton - time: 0,51 GtkProgressBar - time: 0,27 GtkToggleButton - time: 1,34 GtkCheckButton - time: 1,44 GtkRadioButton - time: 1,79 GtkTextView - Add text - time: 11,68 GtkTextView - Scroll - time: 2,84 GtkDrawingArea - Lines - time: 2,43 GtkDrawingArea - Circles - time: 1,86 GtkDrawingArea - Text - time: 3,89 GtkDrawingArea - Pixbufs - time: 0,30 --- Total time: 35,22 1.5.1 system, EXA: GtkEntry - time: 0,16 GtkComboBox - time: 6,52 GtkComboBoxEntry - time: 3,93 GtkSpinButton - time: 0,83 GtkProgressBar - time: 0,40 GtkToggleButton - time: 1,19 GtkCheckButton - time: 0,55 GtkRadioButton - time: 1,45 GtkTextView - Add text - time: 13,90 GtkTextView - Scroll - time: 2,23 GtkDrawingArea - Lines - time: 5,48 GtkDrawingArea - Circles - time: 4,04 GtkDrawingArea - Text - time: 6,94 GtkDrawingArea - Pixbufs - time: 0,36 --- Total time: 47,97 ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, Oct 09, 2008 at 11:41:21AM +0200, Michel Dänzer wrote: http://cgit.freedesktop.org/xorg/xserver/commit/?h=server-1.5-branchid=8ef37c194fa08d3911095299413a42a01162b078 Well, doesn't it need also the patch that converts all DevPrivateKey key=key; to static int keyIndex; DevPrivateKey key=keyIndex ? -- Lukáš Hejtmánek ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, 2008-10-09 at 15:47 +0200, Lukas Hejtmanek wrote: On Thu, Oct 09, 2008 at 11:41:21AM +0200, Michel Dänzer wrote: http://cgit.freedesktop.org/xorg/xserver/commit/?h=server-1.5-branchid=8ef37c194fa08d3911095299413a42a01162b078 Well, doesn't it need also the patch that converts all DevPrivateKey key=key; to static int keyIndex; DevPrivateKey key=keyIndex ? You mean commit 4017d3190234e189a0bbd33193a148d4d3c7556b? That's already in 1.4.99.901. -- Earthling Michel Dänzer | http://tungstengraphics.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, 2008-10-09 at 11:14 +0200, Michel Dänzer wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. Yeah, but I'm unlikely to actually include it in 1.5.2. It looks like it's enough of a change from the 1.5.0 devprivate semantics to be considered an ABI change. - ajax signature.asc Description: This is a digitally signed message part ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 1.4 - 1.5.1 performance regressions
On Thu, 2008-10-09 at 11:35 +0200, Fabio wrote: On Thu, 2008-10-09 at 11:02 +0200, Fabio wrote: I noticed some performance regression moving to a newer system with Xserver 1.5.1 from an 1.4 system with both XAA and EXA. Can you try current Git server-1.5-branch? Adam Jackson just backported the reduction of dixLookupPrivate() overhead, which is used extensively by the EXA core at least. If there are still regressions with EXA, I'd be interested in sysprof/oprofile data separately for each regressed case. I can't update at the moment (I am using prebuilt packages from Ubuntu). Can you point the fix? I don't see it at: http://cgit.freedesktop.org/xorg/xserver/log/?h=server-1.5-branch http://cgit.freedesktop.org/xorg/xserver/commit/?h=server-1.5-branchid=8ef37c194fa08d3911095299413a42a01162b078 -- Earthling Michel Dänzer | http://tungstengraphics.com Libre software enthusiast | Debian, X and DRI developer ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg