[Bug 850794] New: severe performance regression on amd platform
https://bugzilla.novell.com/show_bug.cgi?id=850794 https://bugzilla.novell.com/show_bug.cgi?id=850794#c0 Summary: severe performance regression on amd platform Classification: openSUSE Product: openSUSE 13.1 Version: Final Platform: x86-64 OS/Version: openSUSE 13.1 Status: NEW Severity: Normal Priority: P5 - None Component: Basesystem AssignedTo: bnc-team-screening@forge.provo.novell.com ReportedBy: gindikin@gmail.com QAContact: qa-bugs@suse.de Found By: --- Blocker: --- User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:25.0) Gecko/20100101 Firefox/25.0 With openSUSE 13.1 I observe a severe performance regression as compared to openSUSE 12.3. Specifically, Wolfram Mathematica 9.0.1 benchmark gives 0.70 on 13.1 as opposed to 0.97 on 12.3. My own Mathematica code runs twice as slow. I think this might be related to broken CPU frequency scaling due to transition to systemd, see https://bbs.archlinux.org/viewtopic.php?id=169372 (maybe not). Specs: 2x Opteron 6272, Supermicro H8DGI-F, 128 GB memory. Reproducible: Always Steps to Reproduce: 1. Install openSUSE 13.1 x86_64 on AMD platform 2. Do Wolfram Mathematica benchmark 3. Compare result to benchmark @ openSUSE 12.3 (or earlier) Actual Results: Observe approx. 30% performance regression Expected Results: No significant regression in benchmark results -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c
Xiyuan Liu
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c1
Borislav Petkov
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c2
--- Comment #2 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c3
--- Comment #3 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c4
--- Comment #4 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c5
--- Comment #5 from Borislav Petkov
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c6
--- Comment #6 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c7
--- Comment #7 from Borislav Petkov
1. perf does not understand the option --pre. Is it a misprint?
Oh, right, --pre is not in oS's version of perf probably. You can use the perf in 3.12, just cd into the kernel source dir of the downloaded 3.12, then do $ cd tools/perf/ $ make -j8 $ cd ../../ and then call the perf binary in there. I.e., /tools/perf/perf ...
2. where do I put the file pre-build-kernel.sh?
you can put it in the same directory where the kernel is or anywhere else you want: /tools/perf/perf stat --pre /home/yasha/bin/pre-build-kernel.sh ... would work too. It needs an absolute path so that perf can do system(3) on it. HTH. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c8
--- Comment #8 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c9
--- Comment #9 from Borislav Petkov
Created an attachment (id=568087) --> (http://bugzilla.novell.com/attachment.cgi?id=568087) [details] The output of ./tools/perf/perf stat --repeat 10 -a --sync --pre /root/pre-build-kernel.sh make -s -j64 bzImage
Thank you for your kind explanations. The console output for openSUSE 12.3 is attached to the comment.
Glad you liked them :-) Now I'd need you to do the exact same perf exercise on 13.1 too :-) Thanks. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c10
--- Comment #10 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c11
--- Comment #11 from Borislav Petkov
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c12
--- Comment #12 from Borislav Petkov
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c13
--- Comment #13 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c14
--- Comment #14 from Yasha Gindikin
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c15
--- Comment #15 from Borislav Petkov
Considering Mathematica benchmark: I have made a clean fresh install of openSUSE 13.1, and you know what? The performance drop is gone. Well, almost. Mathematica default benchmark (Benchmark[]) gives 0.94, which is comparable to that of openSUSE 12.3. My own (highly parallelized) code runs 10% slower, but this drop-down is not as dramatic as the one observed previously. So, maybe we should close the bug, should we?
Yes, we will :) Great, so some libs didn't get updated properly or whatever else happened. I must remember this for the next time to ask bug reporters to install afresh. But yeah, normally one can't know what gets and doesn't get overwritten during a distro upgrade (I'm assuming this is what you did the first time).
If there remain questions pertaining to the general performance of openSUSE 13.1 kernel, I am nonetheless eager to help.
Thanks but I can do it here too - basically I have to run the perf record workload on one of my test boxes here on 3.7 and 3.11 or maybe 3.12 and compare. Also have to check where we're spending the most time. Oh well, when there's time. For the record, hunting performance regressions in the kernel can be very error prone, time consuming and tedious business. (in reply to comment #14)
BTW, the perf stat results above both pertained to the fresh clean installations of openSUSE, where Mathematica performance issues were already absent.
Right. Thanks for your time. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c16
--- Comment #16 from Borislav Petkov
https://bugzilla.novell.com/show_bug.cgi?id=850794
https://bugzilla.novell.com/show_bug.cgi?id=850794#c17
Borislav Petkov
participants (1)
-
bugzilla_noreply@novell.com