Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Hi all, I'm reading the bug news in https://bugzilla.novell.com/show_bug.cgi?id=333739. Maybe I'm being too much newbie, I cannot understand why the audit is enabled by default, and what CONFIG_AUDITSYSCALL option means. Any hint? Rob On 10/19/07, Ben Kevan [EMAIL PROTECTED] wrote: On Thursday 18 October 2007 02:25:04 pm nordi wrote: I built myself a new kernel and surprise: I got A LOT faster! Posix 10.0 UTF8 10.3 10.3 (new kernel) == == Dhrystone336 339 338 Whetstone198 204 206 Execl658 576 632 File Copy 1024 535 481 595 File Copy 256455 355 458 File Copy 4096 588 717 827 Pipe Throughput 468 278 408 Context Switch 554 384 567 Process Creat 1000 783 921 Shell Scripts1 873 344 362 Shell Scripts8 894 332 349 System Call 904 334 819 -- -- -- Index Score: 569 397 496 I have appended my kernel config. It is the standard config minus everything that I thought could potentially hurt syscall performance (8 changes all together). Since building a kernel takes quite some time on my machine I haven't checked exactly which change it was. Note that I used UTF-8 again, but Shell Script performance still went up by 17-18 points. That is a 5% speed increase for a shell script, simply because of the new kernel! Quite likely other applications will benefit as well. I think Suse should really look into this issue, getting 5% more performance in your applications is something that everyone would like to have. But so far, I have gotten no replies to my bug report [1]. Btw, if you have a Bugzilla account you can add yourself to the CC: list and get informed about all changes to this bug. Regards nordi [1] https://bugzilla.novell.com/show_bug.cgi?id=333739 I did similar here.. I actually installed a vanilla 2.6.23.. and boy.. the shit screamed like a little kiddie in a haunted house on halloween.. I love it.. Ben -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] -- Are you tired of making software? Play it! (http://www.codesounding.org) -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Roberto Mannai [EMAIL PROTECTED] writes: Hi all, I'm reading the bug news in https://bugzilla.novell.com/show_bug.cgi?id=333739. Maybe I'm being too much newbie, I cannot understand why the audit is enabled by default, and what CONFIG_AUDITSYSCALL option means. Any hint? There're the Audit Quick Start and Linux Audit Framework manuals on the Novell side ( http://www.novell.com/documentation/sles10/ ) for an in-depth treatment. Audit is a security tool - used not only by AppArmor - that allows you to see what programs are doing. Think of a system wide trace mechanism, Andreas -- Andreas Jaeger, Director Platform/openSUSE, [EMAIL PROTECTED] SUSE LINUX Products GmbH, GF: Markus Rex, HRB 16746 (AG Nürnberg) Maxfeldstr. 5, 90409 Nürnberg, Germany GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126 pgpaqND22MaLh.pgp Description: PGP signature
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
I built myself a new kernel and surprise: I got A LOT faster! Posix 10.0 UTF8 10.3 10.3 (new kernel) == == Dhrystone336 339 338 Whetstone198 204 206 Execl658 576 632 File Copy 1024 535 481 595 File Copy 256455 355 458 File Copy 4096 588 717 827 Pipe Throughput 468 278 408 Context Switch 554 384 567 Process Creat 1000 783 921 Shell Scripts1 873 344 362 Shell Scripts8 894 332 349 System Call 904 334 819 -- -- -- Index Score: 569 397 496 I have appended my kernel config. It is the standard config minus everything that I thought could potentially hurt syscall performance (8 changes all together). Since building a kernel takes quite some time on my machine I haven't checked exactly which change it was. Note that I used UTF-8 again, but Shell Script performance still went up by 17-18 points. That is a 5% speed increase for a shell script, simply because of the new kernel! Quite likely other applications will benefit as well. I think Suse should really look into this issue, getting 5% more performance in your applications is something that everyone would like to have. But so far, I have gotten no replies to my bug report [1]. Btw, if you have a Bugzilla account you can add yourself to the CC: list and get informed about all changes to this bug. Regards nordi [1] https://bugzilla.novell.com/show_bug.cgi?id=333739 config_103fast.bz2 Description: BZip2 compressed data
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Thursday 18 October 2007 02:25:04 pm nordi wrote: I built myself a new kernel and surprise: I got A LOT faster! Posix 10.0 UTF8 10.3 10.3 (new kernel) == == Dhrystone336 339 338 Whetstone198 204 206 Execl658 576 632 File Copy 1024 535 481 595 File Copy 256455 355 458 File Copy 4096 588 717 827 Pipe Throughput 468 278 408 Context Switch 554 384 567 Process Creat 1000 783 921 Shell Scripts1 873 344 362 Shell Scripts8 894 332 349 System Call 904 334 819 -- -- -- Index Score: 569 397 496 I have appended my kernel config. It is the standard config minus everything that I thought could potentially hurt syscall performance (8 changes all together). Since building a kernel takes quite some time on my machine I haven't checked exactly which change it was. Note that I used UTF-8 again, but Shell Script performance still went up by 17-18 points. That is a 5% speed increase for a shell script, simply because of the new kernel! Quite likely other applications will benefit as well. I think Suse should really look into this issue, getting 5% more performance in your applications is something that everyone would like to have. But so far, I have gotten no replies to my bug report [1]. Btw, if you have a Bugzilla account you can add yourself to the CC: list and get informed about all changes to this bug. Regards nordi [1] https://bugzilla.novell.com/show_bug.cgi?id=333739 I did similar here.. I actually installed a vanilla 2.6.23.. and boy.. the shit screamed like a little kiddie in a haunted house on halloween.. I love it.. Ben -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
I just see this in an other thread og this list http://kerneltrap.org/node/5411 should this be related to benchmarking? jdd -- http://www.dodin.net -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Hey, nordi wrote: Looking at the syscalls in more depth, I wrote a small and simple program that _only_ does syscalls in a big loop, see attachment. Just uncomment the syscall that you want to benchmark. Maybe it is from interest: http://www.opensolaris.org/os/project/libmicro/ LibMicro is intended to measure the performance of various system and library calls. With that you have nearly all syscalls tested, even with math. analysis (median, stdddev, 99% confidence level) Regards, -- Patrick Kirsch - Quality Assurance Department SUSE Linux Products GmbH GF: Markus Rex, HRB 16746 (AG Nuernberg) -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 14:22:48 nordi wrote: Forgot to mention that we now have bug #333739, see https://bugzilla.novell.com/show_bug.cgi?id=333739 Thanks for that! Maybe the Suse kernel-guys know if this is a bug or a feature. Yeah, I was wondering... is this because of different configure options SuSE is setting? Or because of the kernel itself? Right now I'm installing 10.2 and 10.3, and hopefully 10.0 and 10.1, on a test machine. Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
I wrote:
Suse 10.3, runlevel 2:385.9
Suse 10.3, runlevel 1:756.9
After noticing that the benchmark also runs significantly faster for
root than it runs for a normal user, I started looking at the
environment. It turns out that the LANG variable is all that matters:
$ echo $LANG
de_DE.UTF-8
$ time for i in {1..1000}; do ../pgms/tst.sh; done
real0m22.807s
user0m16.365s
sys 0m6.388s
$ LANG=POSIX
$ time for i in {1..1000}; do ../pgms/tst.sh; done
real0m12.113s
user0m7.092s
sys 0m5.016s
So this benchmark is not really measuring performance, it is measuring
your language settings. Ian, you should modify all tests to use the same
language settings everywhere, because otherwise the results are pure
bogus. And then re-run the benchmarks on 10.2 and 10.3 and we will
hopefully see a performance _increase_ for 10.3 ;)
Regards
nordi
--
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
[opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Where can I get Unixbench? Are these tests done with or without beagle enabled? -- Regards, Aniruddha Please adhere to the OpenSUSE_mailing_list_netiquette http://en.opensuse.org/OpenSUSE_mailing_list_netiquette -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Aniruddha wrote: Where can I get Unixbench? Simply use the link that Ian provided: http://www.hermit.org/Linux/Benchmarking/ At the bottom of the page you will find links to Unixbench. Are these tests done with or without beagle enabled? I don't have beagle installed, so my results are without beagle. Don't know for Ian. Regards nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
I wrote: Ian, you should modify all tests to use the same language settings everywhere, because otherwise the results are pure bogus. The question is: Should we use POSIX or UTF8? If we use POSIX the results are somehow unrealistic, because everyone uses UTF8 nowadays. If we use UTF8, we cannot compare to older systems that do not support it. And then re-run the benchmarks on 10.2 and 10.3 and we will hopefully see a performance _increase_ for 10.3 ;) Hm, my results are not really what I had hoped for. More testing shows that 10.3 still seems to be much slower than 10.0 on my system: Posix 10.0 UTF8 10.3 Posix 10.3 == == Dhrystone 335.6 339.1 326.9 ok Whetstone 198.4 203.5 201.7 ok Execl 658.3 576.3 573.1 -13% File Copy 1024 534.6 481.0 480.9 File Copy 256 455.2 354.5 353.8 File Copy 4096 588.3 717.4 736.2 Pipe Throughput 468.1 277.6 283.3 -40% Context Switch 554.3 384.1 385.4 -31% Process Creat 1000.2 782.7 770.5 -23% Shell Scripts1 873.0 343.8!!!721.0 -17% Shell Scripts8 893.6 331.7!!!724.6 -19% System Call 903.8 333.7 336.7 -63%!!! - - - Index Score:568.9 397.3 450.6 The first two only do calculations and they are ok, some jitter, not more. The last ones (syscall, pipe, switch, create processes) have a lot of kernel involvement and score very low. The shell scripts also make heavy use of pipes, which might explain why they still score much lower for 10.3 than for 10.0, even though LANG=POSIX is used on both systems. Somehow this does not look right. The kernel in 10.3 seems to be _much_ slower than in 10.0. Maybe someone forgot to activate some optimization in the kernel config? Regards nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 15:40:58 nordi wrote: I wrote: Ian, you should modify all tests to use the same language settings everywhere, because otherwise the results are pure bogus. The question is: Should we use POSIX or UTF8? If we use POSIX the results are somehow unrealistic, because everyone uses UTF8 nowadays. If we use UTF8, we cannot compare to older systems that do not support it. And then re-run the benchmarks on 10.2 and 10.3 and we will hopefully see a performance _increase_ for 10.3 ;) Hm, my results are not really what I had hoped for. More testing shows that 10.3 still seems to be much slower than 10.0 on my system: Posix 10.0 UTF8 10.3 Posix 10.3 == == Dhrystone 335.6 339.1 326.9 ok Whetstone 198.4 203.5 201.7 ok Execl 658.3 576.3 573.1 -13% File Copy 1024534.6 481.0 480.9 File Copy 256 455.2 354.5 353.8 File Copy 4096588.3 717.4 736.2 Pipe Throughput 468.1 277.6 283.3 -40% Context Switch554.3 384.1 385.4 -31% Process Creat 1000.2 782.7 770.5 -23% Shell Scripts1873.0 343.8!!!721.0 -17% Shell Scripts8893.6 331.7!!!724.6 -19% System Call 903.8 333.7 336.7 -63%!!! - - - Index Score: 568.9 397.3 450.6 The first two only do calculations and they are ok, some jitter, not more. The last ones (syscall, pipe, switch, create processes) have a lot of kernel involvement and score very low. The shell scripts also make heavy use of pipes, which might explain why they still score much lower for 10.3 than for 10.0, even though LANG=POSIX is used on both systems. Somehow this does not look right. The kernel in 10.3 seems to be _much_ slower than in 10.0. Maybe someone forgot to activate some optimization in the kernel config? I don't have a 10.0 handy for testing. could you try' strace -T pgms/syscall 3 it will run for 3 seconds, and tell you where it spends its time. btw, I'm assuming you're running this on the same hardware for all tests Also, tests that don't do text manipulation, like grep, don't need to be tested with different locales Anders -- Madness takes its toll -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Seems my table got messed up for whatever reason. Lets try again: Posix 10.0 UTF8 10.3 Posix 10.3 == == Dhrystone335.6339.1 326.9 ok Whetstone198.4203.5 201.7 ok Execl658.3576.3 573.1 -13% File Copy 1024 534.6481.0 480.9 File Copy 256455.2354.5 353.8 File Copy 4096 588.3717.4 736.2 Pipe Throughput 468.1277.6 283.3 -40% Context Switch 554.3384.1 385.4 -31% Process Creat 1000.2782.7 770.5 -23% Shell Scripts1 873.0343.8!!!721.0 -17% Shell Scripts8 893.6331.7!!!724.6 -19% System Call 903.8333.7 336.7 -63%!!! -- -- -- Index Score: 568.9397.3 450.6 Regards nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sun, Oct 14, 2007 at 03:40:58PM +0200, nordi wrote: I wrote: Ian, you should modify all tests to use the same language settings everywhere, because otherwise the results are pure bogus. The question is: Should we use POSIX or UTF8? If we use POSIX the results are somehow unrealistic, because everyone uses UTF8 nowadays. If we use UTF8, we cannot compare to older systems that do not support it. And then re-run the benchmarks on 10.2 and 10.3 and we will hopefully see a performance _increase_ for 10.3 ;) Hm, my results are not really what I had hoped for. More testing shows that 10.3 still seems to be much slower than 10.0 on my system: Posix 10.0 UTF8 10.3 Posix 10.3 == == Dhrystone 335.6 339.1 326.9 ok Whetstone 198.4 203.5 201.7 ok Execl 658.3 576.3 573.1 -13% File Copy 1024534.6 481.0 480.9 File Copy 256 455.2 354.5 353.8 File Copy 4096588.3 717.4 736.2 Pipe Throughput 468.1 277.6 283.3 -40% Context Switch554.3 384.1 385.4 -31% Process Creat 1000.2 782.7 770.5 -23% Shell Scripts1873.0 343.8!!!721.0 -17% Shell Scripts8893.6 331.7!!!724.6 -19% System Call 903.8 333.7 336.7 -63%!!! - - - Index Score: 568.9 397.3 450.6 The first two only do calculations and they are ok, some jitter, not more. The last ones (syscall, pipe, switch, create processes) have a lot of kernel involvement and score very low. The shell scripts also make heavy use of pipes, which might explain why they still score much lower for 10.3 than for 10.0, even though LANG=POSIX is used on both systems. Somehow this does not look right. The kernel in 10.3 seems to be _much_ slower than in 10.0. Maybe someone forgot to activate some optimization in the kernel config? Can you run oprofile on them and see where time is wasted? Ciao, Marcus -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Anders Johansson wrote: I don't have a 10.0 handy for testing. could you try' strace -T pgms/syscall 3 it will run for 3 seconds, and tell you where it spends its time. I have attached the trace for Suse Linux 10.0. btw, I'm assuming you're running this on the same hardware for all tests Sure, wouldn't make a lot of sense if I didn't. Same hardware (Pentium M 1.3Ghz), same binaries (as downloaded for unixbench5.0). Also, tests that don't do text manipulation, like grep, don't need to be tested with different locales You are right, just wanted to be _really_ sure. Regards nordi trace_anders.bz2 Description: BZip2 compressed data
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 02:51:40 Aniruddha wrote: Where can I get Unixbench? See my original post: You'll find the benchmarks, system details, and full results at: http://www.hermit.org/Linux/Benchmarking/ Are these tests done with or without beagle enabled? No beagle, I always delete it hey, I know where things are... ;-) Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 16:58:28 nordi wrote: Anders Johansson wrote: I don't have a 10.0 handy for testing. could you try' strace -T pgms/syscall 3 it will run for 3 seconds, and tell you where it spends its time. I have attached the trace for Suse Linux 10.0. Cool. Can you do the same thing for 10.3, so we can compare? Anders -- Madness takes its toll -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
What a response! Thanks everyone. I'm going to consolidate some replies here... On Saturday 13 October 2007 13:09:37 nordi wrote: Suse 10.0, runlevel 5:511.4 Suse 10.0, runlevel 1:920.7 Suse 10.3, runlevel 2:385.9 Suse 10.3, runlevel 1:756.9 This is _very_ strange. Usually I would say the benchmark is broken, but the benchmark simply starts a shell script that starts some GNU utilities. There's not much you can break here. Please note that the benchmarks themselves haven't been touched in 10 years at least (all my work has been on the framework around them). There could be all sorts of weirdness in there. But I don't think so, they're really very simple (too simple, really). Can someone confirm that running in runlevel 1 yields much higher benchmark scores? Yes: as the USAGE file says: When running the tests, I do *not* recommend switching to single-user mode (init 1). This seems to change the results in ways I don't understand, and it's not realistic (unless your system will actually be running in this mode, of course). No idea why, though. On Saturday 13 October 2007 13:43:37 Anders Johansson wrote: But yes, the benchmark is broken. I haven't looked in any great detail at what it does, but how it measures it is just wrong. In theory, it runs for 60 seconds, and at the end it counts how many iterations it has managed to do in that time, averaged over a couple of runs The problem is that it never checks if it has run for 60 seconds. It sets up a signal handler for SIGALRM, and just assumes that when the process receives that signal, the 60 seconds are up and it's time to report. Not true. Most test don't report the time taken; the Run script measures the *actual* time the test runs for, and uses that figure. This is not ideal, because it includes the program start-up and shutdown in the test score, but that's how it's always been. I'm considering changing it, though; I already did for the FS tests. On Saturday 13 October 2007 16:24:53 Lew Wolfgang wrote: I didn't try Ian's benchmarks, but I did fiddle around a bit with a floating-point intensive one that I've been using for years. It calculates very long FFT's and displays the accuracy. Bottom line is I didn't see any significant differences between runlevels 1 and 5. The benchmark ran in 8.7 seconds as measured by time. It did run a bit faster in 10.3 than 10.2. However, this wasn't a fair test since my 10.2 is 32-bit, my 10.3 64-bit on the same computer Then it's meaningless, I'm afraid. I'm seeing a 10-15% speedup just from running the 64-bit version of Linux as opposed to the 32-bit version, which would compensate for the slowdown in 10.3. Compare like with like. See: http://www.hermit.org/Linux/Benchmarking/ My test shows dhrystone *faster* on 10.3, and double-precision whetstone about the same. So it's no surprise that your FP test shows now slowdown. The slowdown I saw was in context switching, shell scripts (dramatically) and system calls (dramatically). OK, next batch... ;-) Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 02:03:37 nordi wrote: After noticing that the benchmark also runs significantly faster for root than it runs for a normal user, I started looking at the environment. It turns out that the LANG variable is all that matters: [ ... ] So this benchmark is not really measuring performance, it is measuring your language settings. Ian, you should modify all tests to use the same language settings everywhere, because otherwise the results are pure bogus. And then re-run the benchmarks on 10.2 and 10.3 and we will hopefully see a performance _increase_ for 10.3 ;) Thanks for this investigation, it certainly looks relevant. BUT: I don't get what you mean. On one hand, I guess you're saying that I should set LANG manually, so that people running UnixBench all around the world will see consistent results. That's obviously a very good idea, and I'll do that. Thanks for the tip! But how does that change *my* results? My Sony and HP test systems are always installed as UK English; the others (they belong to my employer) as US English. So when I see a slowdown between 10.2 and 10.3 on the Sony, and a similar slowdown on the HP, that must be caused by something else. The slowdown seems to be worst in the shell tests and the system call overhead tests. How would LANG affect the latter? I'm going to re-install 10.2 on one of the Dell's boot partitions and do some more testing... now that I know about LANG, I'll take that into account. Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 06:40:58 nordi wrote: I wrote: Ian, you should modify all tests to use the same language settings everywhere, because otherwise the results are pure bogus. The question is: Should we use POSIX or UTF8? If we use POSIX the results are somehow unrealistic, because everyone uses UTF8 nowadays. If we use UTF8, we cannot compare to older systems that do not support it. Anyone else got any feelings on this? Obviously we need to set it to something consistent. My feeling is that I should set it to en_US.UTF-8. Rationale: * Every (modern) install should support en_US.UTF-8. * Like nordi said, benchmarking with settings no-one uses is going to be unrealistic; for example, say there was a machine with hardware UTF support (it could happen) -- then if the tests were run as POSIX they wouldn't show the improvement. The big drawback, as nordi said, is that you lose consistency with pre-Unicode systems. Or do you? It's the old benchmarking problem of what it is that you're trying to measure. If you're measuring kernel performance, then you should always use POSIX, to remove the effect of things like the shell. But if you're measuring system performance -- which is what UnixBench is really designed for -- then you should use the system's default settings, so you measure what the system really does. After all, Unicode systems *really do* go slower than ASCII systems, and the test results should reflect that. Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 17:22:02 Ian Smith wrote: After all, Unicode systems *really do* go slower than ASCII systems, and the test results should reflect that. Not for people whose languages aren't representable in ASCII. For them, an ASCII system would be much slower If you don't take into account missing functionality, you might as well run your benchmark on a machine with no OS at all (like DOS, for example) and declare it the overall winner. Sure it's faster, but what difference does it make if you can't actually use it for anything useful If all you're interested in is winning benchmarks, I can provide you with patched versions of glibc and bash (where most functions are replaced by NOOP), which would beat all your systems hands down Like you said yourself, compare like with like Anders -- Madness takes its toll -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
OK, I may have cracked the runlevel 1 issue. It seems that LANG is not set in runlevel 1. I just ran the following tests, on the Sony system, OpenSuSE 10.3, in runlevel 1: LANG not defined: System Benchmarks Partial Index BASELINE RESULTINDEX Shell Scripts (1 concurrent) 42.4 2815.8664.1 System Call Overhead 15000.0 454455.1303.0 LANG = en_US.UTF8: System Benchmarks Partial Index BASELINE RESULTINDEX Shell Scripts (1 concurrent) 42.4 1416.8334.2 System Call Overhead 15000.0 455891.6303.9 These last results pretty much match what I saw in runlevel 5 on the same machine. Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Anders Johansson schrieb: Cool. Can you do the same thing for 10.3, so we can compare? Here it is. Syscalls run considerably slower in 10.3 than in 10.0. I grepped through the file to see how long the syscalls took: 10.0: 5 usecs: 52792 6 usecs: 79944 7 usecs: 434 weighted average: 5.61 usecs 10.3 5 usecs: 16811 6 usecs: 85497 7 usecs: 4282 weighted average: 5.88 usecs That's only a 5% difference, but probably there are rounding errors hiding somewhere since we have very small numbers: If you go by the size of the files (6.0MB versus 7.5MB) the difference is much clearer. Btw, I saw the following on Suse 10.3: $ head -n 1 trace_anders temp $ time grep -c 'uid' temp 2494 real0m4.745s user0m4.720s sys 0m0.004s $ LANG=POSIX $ time grep -c 'uid' temp 2494 real0m0.004s user0m0.004s sys 0m0.000s Grepping with UTF-8 was super-slow for me, while grepping with LANG=POSIX worked as expected. On Suse 10.0 I can grep the whole file with UTF-8 in just 0.9 seconds, while on Suse 10.3 it takes 4.7 seconds to grep a small fraction. Anyone else seeing this? Regards nordi trace_slow.bz2 Description: BZip2 compressed data
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 08:38:06 Anders Johansson wrote: If all you're interested in is winning benchmarks, I can provide you with patched versions of glibc and bash (where most functions are replaced by NOOP), which would beat all your systems hands down Like you said yourself, compare like with like So... you're agreeing that we should use UTF-8? Seems sensible, anyhow. One question. I can set LANG to en_US.UTF-8, but I would like to have the test report include the language setting to confirm that it is set right. Like, if a system doesn't support en_US.UTF-8 for some reason, I want to know that it's not running a fair test. So how can I tell what the system is really using? For example, if I set LANG to sfsfgsfdg, then locale tells me I'm using sfsfgsfdg, but it actually defaults back to POSIX, and I get the wrong scores again. The command locale -a | grep $LANG should tell me whether the locale is installed, but doesn't, because it reports the name in a different format! Ideas? Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Ian Smith wrote: The command locale -a | grep $LANG should tell me whether the locale is installed, but doesn't, because it reports the name in a different format! Ideas? Try this: $ LANG=foo $ locale -a /dev/null locale: Cannot set LC_CTYPE to default locale: No such file or directory locale: Cannot set LC_MESSAGES to default locale: No such file or directory locale: Cannot set LC_COLLATE to default locale: No such file or directory If you get any messages on stderr, then the locale is not supported. nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 09:11:24 nordi wrote: $ LANG=foo $ locale -a /dev/null locale: Cannot set LC_CTYPE to default locale: No such file or directory locale: Cannot set LC_MESSAGES to default locale: No such file or directory locale: Cannot set LC_COLLATE to default locale: No such file or directory If you get any messages on stderr, then the locale is not supported. Yeah, it's a bit of a hack, and I'm worried how portable that would be (I'm trying to keep this a Unix benchmark, not just Linux), but that may be the best plan. So, UnixBench 5.2 will set LANG, and do its best to report the setting. Any other feature requests while I'm at it? Cheers, Ian -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
nordi wrote: Try this: $ LANG=foo $ locale -a /dev/null locale: Cannot set LC_CTYPE to default locale: No such file or directory locale: Cannot set LC_MESSAGES to default locale: No such file or directory locale: Cannot set LC_COLLATE to default locale: No such file or directory If you get any messages on stderr, then the locale is not supported. Nope, that doesn't work on Solaris. But Solaris formats its output so that your original instruction works: $ locale -a | grep $LANG de_DE.UTF-8 [EMAIL PROTECTED] Looks like Linux and Solaris use different output formats, and who knows what the other Unixes are doing. Regards nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Sunday 14 October 2007 18:01:32 nordi wrote: Anders Johansson schrieb: Cool. Can you do the same thing for 10.3, so we can compare? Here it is. Syscalls run considerably slower in 10.3 than in 10.0. I grepped through the file to see how long the syscalls took: 10.0: 5 usecs: 52792 6 usecs: 79944 7 usecs: 434 weighted average: 5.61 usecs 10.3 5 usecs: 16811 6 usecs: 85497 7 usecs: 4282 weighted average: 5.88 usecs That's only a 5% difference, but probably there are rounding errors hiding somewhere since we have very small numbers: If you go by the size of the files (6.0MB versus 7.5MB) the difference is much clearer. I think the main difference here is in the time the execve() call takes. On 10.0 it's 0.000140 seconds, whereas on 10.3 it's 0.019752. That's 140 times slower, and it dwarves all the other times I wonder why that would be Anders -- Madness takes its toll -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Looking at the syscalls in more depth, I wrote a small and simple
program that _only_ does syscalls in a big loop, see attachment. Just
uncomment the syscall that you want to benchmark.
Here are the results for Suse 10.0, 10.2 (rescue system) and 10.3.
Hopefully I get the table right this time...
10.0 10.210.3
-- -
gethostname8.5711.77 14.47 seconds/run
stat 14.4919,38 21.90 seconds/run
getuid 2.78 5.438.41 seconds/run
close(dup) 9.0915.83 21.93 seconds/run
Looks like syscalls have been getting slower over time. I'm just amazed
_how_ much slower this is. Did the same thing happen for the vanilla kernel?
Regards
nordi
#include stdio.h
#include unistd.h
#include sys/utsname.h
#include sys/types.h
#include sys/stat.h
int main() {
int i;
uid_t myuid;
struct stat x;
char name[5];
for(i=0;i2000;i++) {
//gethostname(name, 1);
//stat(/, x);
//myuid=getuid();
close(dup(0));
}
return 0;
}
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Forgot to mention that we now have bug #333739, see https://bugzilla.novell.com/show_bug.cgi?id=333739 Maybe the Suse kernel-guys know if this is a bug or a feature. nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
[opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Hi, I've been doing some benchmarking of OpenSUSE using UnixBench 5.1. I noticed that 10.3 is 15% - 25% slower than 10.2. (10.2 was 50% faster than 10.1, yay!). I was wondering if anyone knows why this might be? And is benchmarking a part of the release testing process? You'll find the benchmarks, system details, and full results at: http://www.hermit.org/Linux/Benchmarking/ Cheers, Ian Smith -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Ian Smith wrote: I've been doing some benchmarking of OpenSUSE using UnixBench 5.1. I noticed that 10.3 is 15% - 25% slower than 10.2. (10.2 was 50% faster than 10.1, yay!). The benchmark is really showing very strange numbers. The shell script benchmark (consisting mainly of sort and grep) is only _half_ as fast in 10.3 as in 10.2 in your measurements. I ran that test on my system and got similar results. Interestingly, the performance is much higher if I switch to runlevel 1!!! Here are my results for ./Run shell1 on a Pentium M at 1.3Ghz: Suse 10.0, runlevel 5:511.4 Suse 10.0, runlevel 1:920.7 Suse 10.3, runlevel 2:385.9 Suse 10.3, runlevel 1:756.9 Please note that Ian's Intel Core Duo Processor at 2Ghz scored only 557.9 points on this test, while my much older 1.3Ghz Pentium M scores 756.9 points, at least when I benchmark in runlevel 1. This is _very_ strange. Usually I would say the benchmark is broken, but the benchmark simply starts a shell script that starts some GNU utilities. There's not much you can break here. Can someone confirm that running in runlevel 1 yields much higher benchmark scores? Puzzled nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Saturday 13 October 2007 22:09:37 nordi wrote: Ian Smith wrote: I've been doing some benchmarking of OpenSUSE using UnixBench 5.1. I noticed that 10.3 is 15% - 25% slower than 10.2. (10.2 was 50% faster than 10.1, yay!). The benchmark is really showing very strange numbers. The shell script benchmark (consisting mainly of sort and grep) is only _half_ as fast in 10.3 as in 10.2 in your measurements. I ran that test on my system and got similar results. Interestingly, the performance is much higher if I switch to runlevel 1!!! Here are my results for ./Run shell1 on a Pentium M at 1.3Ghz: Suse 10.0, runlevel 5:511.4 Suse 10.0, runlevel 1:920.7 Suse 10.3, runlevel 2:385.9 Suse 10.3, runlevel 1:756.9 Please note that Ian's Intel Core Duo Processor at 2Ghz scored only 557.9 points on this test, while my much older 1.3Ghz Pentium M scores 756.9 points, at least when I benchmark in runlevel 1. This is _very_ strange. Usually I would say the benchmark is broken, but the benchmark simply starts a shell script that starts some GNU utilities. There's not much you can break here. Can someone confirm that running in runlevel 1 yields much higher benchmark scores? Well, runlevel 1 has nothing running, so anything you do will have the machine more or less to itself. I would be surprised if you didn't get higher scores there But yes, the benchmark is broken. I haven't looked in any great detail at what it does, but how it measures it is just wrong. In theory, it runs for 60 seconds, and at the end it counts how many iterations it has managed to do in that time, averaged over a couple of runs The problem is that it never checks if it has run for 60 seconds. It sets up a signal handler for SIGALRM, and just assumes that when the process receives that signal, the 60 seconds are up and it's time to report. This isn't a good idea for any number of reasons Now let's see what the actual job does. Anders -- Madness takes its toll -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
Anders Johansson wrote: Well, runlevel 1 has nothing running, so anything you do will have the machine more or less to itself. I would be surprised if you didn't get higher scores there Higher scores, yes. But I would expect 1% more, maybe even 5% more. But certainly not 100% higher performance. Otherwise that would mean the few services running in runlevel 2 eat up 50% of my CPU time. I also tried killing services in runlevel 2, hoping to reach the performance of runlevel 1. I killed every demon I could find, unloaded some kernel modules... nothing changed. Since you said the benchmark uses signals for its timing: Is there anything that might send signals in one runlevel and not send them in another? Regards nordi -- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
On Saturday 13 October 2007 22:43:37 Anders Johansson wrote:
Now let's see what the actual job does.
Not sure if I'm missing something here, but...
As far as I can see, the shell1 test simply runs pgms/tst.sh over and over
again, for 60 second
I run it, and I got a score of 316 for the 1 concurrent instance.
Then I ran time for i in {1..700}; do ../pgms/tst.sh; done and it took
21.179 seconds
Somewhere, there is something I'm missing
Anders
--
Madness takes its toll
--
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
Re: [opensuse] OpenSUSE 10.3 benchmarking: slower than 10.2?
nordi wrote:
Ian Smith wrote:
I've been doing some benchmarking of OpenSUSE using UnixBench 5.1. I
noticed that 10.3 is 15% - 25% slower than 10.2. (10.2 was 50% faster
than 10.1, yay!).
The benchmark is really showing very strange numbers. The shell script
benchmark (consisting mainly of sort and grep) is only _half_ as fast in
10.3 as in 10.2 in your measurements. I ran that test on my system and
got similar results. Interestingly, the performance is much higher if I
switch to runlevel 1!!! Here are my results for ./Run shell1 on a
Pentium M at 1.3Ghz:
I didn't try Ian's benchmarks, but I did fiddle around
a bit with a floating-point intensive one that I've
been using for years. It calculates very long FFT's
and displays the accuracy.
Bottom line is I didn't see any significant differences
between runlevels 1 and 5. The benchmark ran in 8.7
seconds as measured by time.
It did run a bit faster in 10.3 than 10.2. However, this
wasn't a fair test since my 10.2 is 32-bit, my 10.3 64-bit
on the same computer. Interesting nevertheless, and all
in runlevel 5.
10.2 (with 10.2 binary) 0m9.866s
10.3 (with 10.2 binary) 0m9.871s
10.3 (with 10.3 binary) 0m8.734s
Here are the particulars:
10.2
Linux train 2.6.18.8-0.5-bigsmp #1 SMP Fri Jun 22 12:17:53 UTC 2007 i686 i686
i386 GNU/Linux
10.3
Linux train 2.6.22.9-0.4-default #1 SMP 2007/10/05 21:32:04 UTC x86_64 x86_64
x86_64 GNU/Linux
10.2
real0m9.866s
user0m9.825s
sys 0m0.040s
10.3 (running 10.2 binary)
real0m9.871s
user0m9.813s
sys 0m0.060s
10.3 (running 10.3 binary)
real0m8.734s
user0m8.701s
sys 0m0.036s
Here's the source for the benchmark. For the purposes
of this report, all versions were compiled:
gcc -O3 edelbench.c -lm
/* Dave Edelblut's Benchmark */
#include math.h
#include time.h
#define NP_MAX 8388608
/* #define NP_MAX 16777216 */
#define DATA_TYPE float
/* #define DATA_TYPE double */
typedef struct { DATA_TYPE r; DATA_TYPE i; } complex;
/*
A Duhamel-Hollman split-radix dif fft
Ref: Electronics Letters, Jan. 5, 1984
Complex input and output data in arrays x and y
Length is n
*/
int cfft( complex *x, int np )
{
int i,j,k,m,n,i0,i1,i2,i3,is,id,n1,n2,n4 ;
DATA_TYPE a,e,a3,cc1,ss1,cc3,ss3,r1,r2,s1,s2,s3,xt ;
x = x - 1;
i = 2; m = 1; while (i np) { i = i+i; m = m+1; };
n = i; if (n != np) {
for (i = np+1; i = n; i++) { x[i].r=0.0; x[i].i=0.0; };
/* printf(\nuse %d point fft\n,n); */ }
n2 = n+n;
for (k = 1; k = m-1; k++ ) {
n2 = n2 / 2; n4 = n2 / 4; e = 2.0 * M_PI / n2; a = 0.0;
for (j = 1; j= n4 ; j++) {
a3 = 3.0*a; cc1 = cos(a); ss1 = sin(a);
cc3 = cos(a3); ss3 = sin(a3); a = j*e; is = j; id = 2*n2;
while ( is n ) {
for (i0 = is; i0 = n-1; i0 = i0 + id) {
i1 = i0 + n4; i2 = i1 + n4; i3 = i2 + n4;
r1= x[i0].r - x[i2].r;
x[i0].r = x[i0].r + x[i2].r;
r2= x[i1].r - x[i3].r;
x[i1].r = x[i1].r + x[i3].r;
s1= x[i0].i - x[i2].i;
x[i0].i = x[i0].i + x[i2].i;
s2= x[i1].i - x[i3].i;
x[i1].i = x[i1].i + x[i3].i;
s3= r1 - s2; r1 = r1 + s2; s2 = r2 - s1; r2 = r2 + s1;
x[i2].r = r1*cc1 - s2*ss1;
x[i2].i = -s2*cc1 - r1*ss1;
x[i3].r = s3*cc3 + r2*ss3;
x[i3].i = r2*cc3 - s3*ss3;
}
is = 2*id - n2 + j; id = 4*id;
}
}
}
/*
-Last stage, length=2 butterfly-
*/
is = 1; id = 4;
while ( is n) {
for (i0 = is; i0 = n; i0 = i0 + id) {
i1 = i0 + 1; r1 = x[i0].r;
x[i0].r = r1 + x[i1].r;
x[i1].r = r1 - x[i1].r;
r1 = x[i0].i;
x[i0].i = r1 + x[i1].i;
x[i1].i = r1 - x[i1].i;
}
is = 2*id - 1; id = 4 * id; }
/*
c--Bit reverse counter
*/
j = 1; n1 = n - 1;
for (i = 1; i = n1; i++) {
if (i j) {
xt = x[j].r;
x[j].r = x[i].r; x[i].r = xt;
xt = x[j].i; x[j].i = x[i].i;
x[i].i = xt;
}
k = n / 2; while (k j) { j = j - k; k = k / 2; }
j = j + k;
}
return(n);
}
/*
program to test fast fourier transform in double precision;
*/
void main()
{
int i,j,ib,np,npm,n2,kr,ki;
double a,enp,t,rx,y,zr,zi,pi,el_t;
clock_t ct0,ct1,ct2,ctd;
static complex x[NP_MAX];
pi = M_PI;
np = 1024;
ct0 = clock();
printf(\n fft benchmark - double precision - GNU C\n);
while (np = NP_MAX){
printf(np =%7d,np); enp = np; npm = np/2-1; t = pi/enp;
x[0].r = (enp - 1.0) / 2.0; x[0].i = 0;
n2 = np / 2; x[n2].r = -0.5; x[n2].i = 0.0;
for (i = 1; i = npm; i++) { j = np - i;
x[i].r = -0.5; x[j].r = -0.5;
y = t * i; y = -cos(y)/sin(y)/2.0;
x[i].i = y; x[j].i = -y;
}
ct1 = clock(); i = cfft(x,np); ct2 = clock(); ctd = ct2 - ct1;
el_t = (double) ctd;
