| Summary: | cairo's performance drops greatly caused by the kernel | ||||||
|---|---|---|---|---|---|---|---|
| Product: | DRI | Reporter: | zhao jian <jian.j.zhao> | ||||
| Component: | DRM/Intel | Assignee: | Carl Worth <cworth> | ||||
| Status: | CLOSED WORKSFORME | QA Contact: | |||||
| Severity: | normal | ||||||
| Priority: | high | Keywords: | regression | ||||
| Version: | XOrg git | ||||||
| Hardware: | Other | ||||||
| OS: | Linux (All) | ||||||
| Whiteboard: | |||||||
| i915 platform: | i915 features: | ||||||
| Attachments: |
|
||||||
(In reply to comment #0) > System Environment: > ---------------------- > Platform: G41 > Arch: x86_64 > OSD: Fedora release 9 (Sulphur) > Libdrm: (master)5a73f066ba149816cc0fc2de4b97ec4714cf8ebc > Mesa: (master)03607708b0499816291f0fb0d1c331fbf034f0ba > Xserver: (master)a85523dc50f392a33a1c00302a0946828bc9249d > Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 > Kernel: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 Thanks for the bug report. The details above showing the versions at which the regression first appeared are very appreciated. Thanks! What's missing is the previously tested versions at which things were last seen to be working. From a separate report, I believe these are the working versions: Last known versions without regression -------------------------------------- Libdrm: (master)30449829c0347dc7dbe29acb13e49e2f2cb72ae9 Mesa: (master)506bacb8e40b0a170a4b620113506925d2333735 Xserver: (master)b1c3dc6ae226db178420e3b5f297b94afc87c94c Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 Kernel_unstable: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 Let me know if I didn't get those right. -Carl (In reply to comment #1) > (In reply to comment #0) > > System Environment: > > ---------------------- > > Platform: G41 > > Arch: x86_64 > > OSD: Fedora release 9 (Sulphur) > > Libdrm: (master)5a73f066ba149816cc0fc2de4b97ec4714cf8ebc > > Mesa: (master)03607708b0499816291f0fb0d1c331fbf034f0ba > > Xserver: (master)a85523dc50f392a33a1c00302a0946828bc9249d > > Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 > > Kernel: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 > Thanks for the bug report. > The details above showing the versions at which the regression > first appeared are very appreciated. Thanks! > What's missing is the previously tested versions at which things were > last seen to be working. From a separate report, I believe these are > the working versions: > Last known versions without regression > -------------------------------------- > Libdrm: (master)30449829c0347dc7dbe29acb13e49e2f2cb72ae9 > Mesa: (master)506bacb8e40b0a170a4b620113506925d2333735 > Xserver: (master)b1c3dc6ae226db178420e3b5f297b94afc87c94c > Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 > Kernel_unstable: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 > Let me know if I didn't get those right. > -Carl No. Carl, I think you have some misunderstanding. As is said in my bug description, I first found this regression with the code of 20090729 compared with our Q2 release. And finally I tested with the code of 20090729 and only change the kernel to qa-branch(d2a55f12b2425ed1febb9e29e10bb405b91aa8dd) it works well, so I think it is caused by kernel. You can just try with the commit of 20090729 and compare drm-intel-next and qa-branch. And the code of 20090729 is: Libdrm: (master)5a73f066ba149816cc0fc2de4b97ec4714cf8ebc Mesa: (master)03607708b0499816291f0fb0d1c331fbf034f0ba Xserver: (master)a85523dc50f392a33a1c00302a0946828bc9249d Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 Kernel: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 (bad) Kernel: (qa-branch) d2a55f12b2425ed1febb9e29e10bb405b91aa8dd (good) Sigh, it's only 2 million lines of changes to review... My fear, given the nature of the traces, is that it is some subtle interaction of the memory manager. I will have a look and see if I can reproduce this with the current stack on top of those two kernels on my g45. I am loathe to write off a 25% performance hit without understanding it, but we may never find the cause. :( Using benchmark/poppler.trace on g45: [2.6.30.1] d2a55f12b2425ed1febb9e29e10bb405b91aa8dd - 8.640s [drm-intel-next, 2.6.35-rc4+] d44a78e83f7549b3c4ae611e667a0db265cf2e00 - 8.460 [fair-eviction, 2.6.35-rc5+] - 9.040s So I can not find evidence to support the argument that the mysterious slowdown is still present in current kernels. Though on this particular setup moving to a LRU fair-eviction scheme to avoid the page-fault-of-doom livelock will hurt. Even more bizarrely, I find the behaviour was reversed: [2.6.31-rc2, bad!] 2a2430f4542467502d39660bfd66b0004fd8d6a - 7.929s Lets investigate a bit further. However, I've hit a kernel panic in the midst of the bisect with only 2268 revisions left to test. Not fun. So for completeness the cause of the speed-up on my system was: commit 52dc7d32b88156248167864f77a9026abe27b432 Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Sat Jun 6 09:46:01 2009 +0100 drm/i915: Clear fence register on tiling stride change. The fence register value also depends upon the stride of the object, so we need to clear the fence if that is changed as well. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> [anholt: Added 8xx and 965 paths, and renamed the confusing i915_gem_object_tiling_ok function to i915_gem_object_fence_offset_ok] Signed-off-by: Eric Anholt <eric@anholt.net> Quite puzzling. Does anyone volunteer to bisect between 52dc7d and 2.6.35 to find the lost perfomance? (In reply to comment #7) > So for completeness the cause of the speed-up on my system was: > commit 52dc7d32b88156248167864f77a9026abe27b432 > Author: Chris Wilson <chris@chris-wilson.co.uk> > Date: Sat Jun 6 09:46:01 2009 +0100 > drm/i915: Clear fence register on tiling stride change. > The fence register value also depends upon the stride of the object, so we > need to clear the fence if that is changed as well. > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > [anholt: Added 8xx and 965 paths, and renamed the confusing > i915_gem_object_tiling_ok function to i915_gem_object_fence_offset_ok] > Signed-off-by: Eric Anholt <eric@anholt.net> > Quite puzzling. Does anyone volunteer to bisect between 52dc7d and 2.6.35 to > find the lost perfomance? I tested with poppler.trace on G41, with kernel 2.6.34 it is 9.302s, with 2.6.31 it is 8.533s and with 2.6.32 it is 9.732. So I think it may be regressed between 2.6.31 and 2.6.32. And I will do more test tomorrow. Closing verified+fixed. No activity on ~6 years. |
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.
Created attachment 28415 [details] xorg.0.log System Environment: ---------------------- Platform: G41 Arch: x86_64 OSD: Fedora release 9 (Sulphur) Libdrm: (master)5a73f066ba149816cc0fc2de4b97ec4714cf8ebc Mesa: (master)03607708b0499816291f0fb0d1c331fbf034f0ba Xserver: (master)a85523dc50f392a33a1c00302a0946828bc9249d Xf86_video_intel: (master)50e2a6734de43a135aa91cd6e6fb5147e15ce315 Kernel: (drm-intel-next)2a2430f4542467502d39660bfd66b0004fd8d6a9 Bug Description: --------------------- I test with cairo-perf on G41, find there is regression when test with poppler-alt-20090608.trace and poppler-20090608.trace. Maybe they are the same issue. The performance data I get with the code of 20090729 is 25% slower than the data with our Q2 release code. And I find the regression is caused by the kernel, if I only change the kernel from drm-intel-next to qa branch(d2a55f12b2425ed1febb9e29e10bb405b91aa8dd),it performs much better. with code of 20090729 and kernel on drm-intel-next(2a2430f4542467502d39660bfd66b0004fd8d6a9): poppler-alt-20090608.trace [ # ] backend test min(s) median(s) stddev. count [ 0] image poppler-alt-20090608 47.803 47.920 0.24% 6/6 [ 0] xlib poppler-alt-20090608 253.262 253.386 0.25% 6/6 poppler-20090608.trace [ # ] backend test min(s) median(s) stddev. count [ 0] image poppler-20090608 47.770 48.047 0.24% 6/6 [ 0] xlib poppler-20090608 252.956 255.086 0.46% 6/6 if I still use code of 20090729 and only change the kernel to qa branch (d2a55f12b2425ed1febb9e29e10bb405b91aa8dd): poppler-alt-20090608.trace.KMS [ # ] backend test min(s) median(s) stddev. count [ 0] image poppler-alt-20090608 46.875 47.566 0.95% 6/6 [ 0] xlib poppler-alt-20090608 196.346 197.292 0.41% 6/6 poppler-20090608.trace.KMS [ # ] backend test min(s) median(s) stddev. count [ 0] image poppler-20090608 46.587 46.744 0.61% 6/6 [ 0] xlib poppler-20090608 196.546 197.130 0.29% 6/6 Reproduce Steps: --------------------- 1. xinit& 2. cairo-perf-trace poppler-alt-20090608.trace(poppler-20090608.trace)