Announcement

Collapse
No announcement yet.

Intel Linux Graphics On Ubuntu Still Flaky

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #21
    Thanks, Michael Larabel for continuing to post meaningless benchmarks and drawing wrong conclusions.

    Originally posted by pvtcupcakes View Post
    From my experience on Arch with Linux 2.6.30, I get 850fps on glxgears with 2.7 and 550 with 2.8.
    There is a technical reason of why 2.7 (I assume EXA/DRI1) performs faster in glxgears than 2.8 (UXA/DR2):

    Originally posted by Keith Packard
    The difference between DRI1 and DRI2 is due in part to the context switch necessary to get buffer swap commands from the DRI2 application to the X server which owns the ?real? front buffer. For an application like glxgears which draws almost nothing, and spends most of its time clearing and swapping, the impact can be significant (note, glxgears is not a benchmark, this is just one of many reasons). On the other hand, having private back buffers means that partially obscured applications will draw faster, not having to loop over clip rectangles in the main rendering loop.


    But as Keith already mentions: glxgears is not a benchmark. But maybe you need one more:

    Originally posted by Carl Worth
    Nobody measures the performance of "draw the same triangles over and over". And if someone does, (by seriously quoting glxgear fps numbers, for example), then everybody gets a good laugh. In fact, the phrase "glxgears is not a benchmark" is a catchphrase among 3D developers.


    Now that we laughed about you and Michael for drawing conclusions from this same benchmark (just ported to QT4):
    Instead of highly popular pictures of llamas today I'll post a few numbers. Not related to llamas at all. Zero llamas. These will be Qt/KDE ...


    Let's consider some serious aspects. We need real application benchmarks such as game engines, firefox rendering or compiz performance. But even in this case, performance may degrade from one version to another. If that's the reason one should check whether one of these reasons apply:

    . I noticed in Urban Terror, when performing Mesa benchmarks, that in certain revisions of the Mesa Stack on x3100 GM965 hardware, some visual effects were not drawn e.g. some lightning or the shot of your weapon. If it's not drawn the entire scene draws faster (though with slight visual corruption or flaws) and you get higher fps reports.

    . In some of the recent versions of the intel driver tearing disappeared. I'm not aware of in which version exactly and whether it applies only to indirect or direct rendering, but this might may a dfficerence. I remember Jesse talking about double buffering:

    I don't know whether this entered mainline or whether it's related. BUT: I don't see tearing with 2.8.0 in compiz enabled X-Org anymore and I know that double buffering has an impact on reported framerates (compared to tearing withough v-sync):


    --

    Michael, please remove QGears2 benchmarks from the benchmark suite that you use for performance evaluation. That applies to the GTK benchmarks as well that you used in this article, which I critized already some time ago:

    Comment


    • #22
      Originally posted by 7oby View Post
      Thanks, Michael Larabel for continuing to post meaningless benchmarks and drawing wrong conclusions.
      Out of interest, how do you suggest you provide a quantative answer to the "Is Ubuntu 9.10 looking to be faster or slower than 9.04 for Intel Users?".

      Comment


      • #23
        Originally posted by mtippett View Post
        Out of interest, how do you suggest you provide a quantative answer to the "Is Ubuntu 9.10 looking to be faster or slower than 9.04 for Intel Users?".
        Why does this question need to be answered on an alpha? (From a publication stand point... I can certainly see it's value from a build quality perspective, but this should be more automated and focused on identifying bottlenecks and regressions, not a "review")

        Comment


        • #24
          Originally posted by Craig73 View Post
          Why does this question need to be answered on an alpha? (From a publication stand point... I can certainly see it's value from a build quality perspective, but this should be more automated and focused on identifying bottlenecks and regressions, not a "review")
          The alpha already contains most of the components that will make it to the final build. In other words, the performance of the alphas should be indicative of the performance of the final build, unless some specific regression is identified and fixed in the meanwhile (and those tests *do* help in finding regressions).

          This was the case with 9.04 for example: the final build had pretty much the same performance as the alphas, at least on my hardware (Ati R500 with the OSS drivers).

          Comment


          • #25
            Originally posted by BlackStar View Post
            The alpha already contains most of the components that will make it to the final build. In other words, the performance of the alphas should be indicative of the performance of the final build, unless some specific regression is identified and fixed in the meanwhile (and those tests *do* help in finding regressions).

            This was the case with 9.04 for example: the final build had pretty much the same performance as the alphas, at least on my hardware (Ati R500 with the OSS drivers).
            I had pretty much the opposite experience with the 8.10->9.04 upgrade. The Phoronix review of Intel performance on the 9.04 Alpha indicated a performance regression, but I saw 3x the 3D performance on my G43-based motherboard.

            Comment


            • #26
              Is ubuntu 9.10 going to be faster then 9.04 for intel users?

              You answer: We tested it with phoronix-test-suite and our conclusion is that the performance will be the same but slightly more unstable. But we tested the newest intel drivers and rc kernel on another distro (like arch, gentoo etc) and it showed boost in performance so we can't be sure at this point.

              I untill recently was very dissapointed in intel's performance although it has advantages over other gpu's because I can built it straight into kernel and I don't have to worry about will nVIDIA or ATI release a driver that will compile on this rc kernel. I like oss and that is one of many reasons I use linux. But now, when they finally did a decent job I read more dissapointing news here on phoronix... Maybie those guys @canonical are doing something wrong but as I say 2.6.31+2.8.0 works perfectly stable and as fast as intel drivers ever where...

              Comment


              • #27
                Originally posted by Craig73 View Post
                Why does this question need to be answered on an alpha? (From a publication stand point... I can certainly see it's value from a build quality perspective, but this should be more automated and focused on identifying bottlenecks and regressions, not a "review")
                It's called performance management. You maintain visibility on performance during development so that you can track when and if there are performance issues and deal with them before production. Realistically, you do not see any public reference to performance of Ubuntu on an ongoing basis except on Phoronix. I trust that as Ubuntu 9.10 gets closer to release, Michael will extend the testing to include the history and ups and downs as it gets closer.

                In general, if you look at some of the "this made some one look really bad" articles, the group with the gap go away, do a few blog postings and discuss what happened. This happened with the SuSE barrier usage. It caused discussion, it caused awareness.

                Phoronix is just taking a mid-stream alpha and benchmarking it.

                If the numbers are wrong, do a PTS response to show where the numbers are wrong.

                If the numbers are right, work with the Ubuntu community to close the gap, offer to help Bryce select the right patch set for the intel drivers. You only have a few milestones left.

                If you don't believe the tests are valid, help define some valid ones to improve the quality of the information.

                Matt

                Comment


                • #28
                  Ok here are my results:

                  gtkperf: 2.6.30 - 1676 ; 2.6.31-rc4 - 1573
                  glmark: 2.6.30 - 64 ; 2.6.31-rc4 - 70
                  urban terror: 2.6.30 - 20.35 ; 2.6.31-rc4 - 23.70

                  So I don't think that there is regression on 2.6.31-rc just an improvement

                  Comment


                  • #29
                    Originally posted by mtippett View Post
                    Out of interest, how do you suggest you provide a quantative answer to the "Is Ubuntu 9.10 looking to be faster or slower than 9.04 for Intel Users?".
                    I answered that already in my previous posting: We need real application benchmarks such as game engines, firefox rendering or compiz performance.

                    It's not a question of the phoronix test suite itself, but of the benchmarks choosen. If developers itself need profiling software to recognize which software pieces are used and to detect software hot spots, what do you think what kind of tools a reviewer needs?

                    Well, Eric Anholt needs cairo-perf-trace:
                    Thanks to cairo-perf-trace, I've just landed a 10% improvement in firefox performance for UXA
                    Another quarter, and another release. I think we've made big progress here. One of my favorite reports on the mailing list was that a company deploying our graphics driver was delightfully surprised that their XV tearing issues were fixed. That was a lot of work, and I was despairing of them saying…


                    Phoronix needs to recognize that it's low level synthetic benchmarks are meaningless. It needs to focus on real applications, because it's simply not experienced enough to work with other benchmarks in order to draw the right conclusions.

                    Comment


                    • #30
                      Originally posted by 7oby View Post
                      I answered that already in my previous posting: We need real application benchmarks such as game engines, firefox rendering or compiz performance.

                      It's not a question of the phoronix test suite itself, but of the benchmarks choosen. If developers itself need profiling software to recognize which software pieces are used and to detect software hot spots, what do you think what kind of tools a reviewer needs?

                      Well, Eric Anholt needs cairo-perf-trace:

                      Another quarter, and another release. I think we've made big progress here. One of my favorite reports on the mailing list was that a company deploying our graphics driver was delightfully surprised that their XV tearing issues were fixed. That was a lot of work, and I was despairing of them saying…
                      Profiling is for developers, benchmarking is for users.

                      Benchmarking determines a macro issue that needs attention, profiling allows developers to deep dive into hotspots that need attention and provide improvements in the benchmark.

                      Note that cairo-perf-trace generates an application call profile, which then can be played as a benchmark. The developers then profile the benchmark to get the hotspots.

                      Phoronix needs to recognize that it's low level synthetic benchmarks are meaningless. It needs to focus on real applications, because it's simply not experienced enough to work with other benchmarks in order to draw the right conclusions.
                      As has been mentioned before, if people are concerned about the quality or correctness of the benchmarks themselves, by all means, invest the effort into getting cairo-perf-trace into PTS. Be part of the solution .

                      A final comment...

                      I don't think Phoronix pulls many conclusions other than the benchmarks presented show a performance delta. What does happen though is the parties that are implicated by the performance issue actually begin to communicate, an example is



                      Bryce (Canonical) concurred about a potential performance issue associated with the intel drivers. He raises a bug to track, intel acknowledges and looks to reproduce the issue. The critical part is that a DISCUSSION is now underway to quantify if there is an issue.

                      That is the value of the Phoronix benchmark articles. It raises to the surface potential issues and triggers a discussion that would most likely not have occured in the past.

                      Regards,

                      Matthew

                      Comment

                      Working...
                      X