Early Linux 2.6.31 Kernel Benchmarks

[benow]

Disk transactions

Why are disk transactions so much faster with 2.6.30+ vs 2.6.29? To me, it seems this would lead to a huge system speedup, no? Is the poor 2.6.29 transaction performance related to the recent sqlite performace problems?

[thefirstm]

I didn't expect to gain much speed by recompiling. My primary objective was to make the system boot faster, which I accomplished. By the way, I also changed from "Voluntary Preemption" to "Preemption" which makes my system respond better when under load. (So well, in fact, that sometimes I have trouble telling it is under load without looking at top.) And having a shorter compile time is nice, too. A stock Ubuntu kernel takes almost an hour to compile on my machine, if I remember correctly. Mine takes only 20 or 30 minutes. I know my startup is slow, it is because I have removed the default Plasma Widget Network Manager because it does not work. Whenever I do this, the system gets some configuration problem, and refuses to start quickly anymore.

[Kano]

You gain only compile time when you disable unused modules and a bit hd space. Your pc will not get faster that way. The different cpu optimization gives you basically no speed increase, maybe in some rare cases. You "feel" changes like a different sheduler but no optimisation settings. Btw. KDE 4 starts really slow, when you would test my own distro with KDE 3.5 as development version then you would see that. Kubuntu is very easy to beat in bootup time.

[thefirstm]

Nice, did you configure that by hand & trial and error or did you use some kind of a tool?

By hand, every bit of it :-) I started with the Ubuntu kernel configuration, then I changed the CPU settings to fit my system better (optimized for Core 2/later Xeon, SMT support off, AMD support off, etc) and then I went through set the drivers for all of my hardware to be compiled in, and I disabled a bunch of stuff I thought I would never use. It booted up the first time, but there were a few more modules I compiled in after that.

[hax0r]

I am using a custom-built image of 2.6.31-rc2 on Kubuntu Karmic alpha. With all of the drivers for the hardware I need compiled in...

Nice, did you configure that by hand & trial and error or did you use some kind of a tool?

Hmm I might try Unigine Tropics / UT2004 on kernels 2.6.20~2.6.24~2.6.29~2.6.30.1~git linux-2.6 and test some 3D.

[thefirstm]

I am using a custom-built image of 2.6.31-rc2 on Kubuntu Karmic alpha. With all of the drivers for the hardware I need compiled in (Literally, the only thing in the output of lsmod is nvidia) and all the stuff I dont need removed, it takes about 11 seconds to boot from scratch to KDM. Then, it takes another 30 or 35 to start KDE. (Somehow, it starts taking this long whenever I modify or remove the Network Manager plasma widget after installing.)

[Kano]

Well pure rc2 is not that good, you need extra patches to make dm work or to compile with older gcc. Also vbox 3 as host has issues on some systems (somehow not all). I did not notice speed diverences yet as i did not do any benchmarks. Basically it is possible to build fglrx with patches without modifiying the kernel too, i just don't know if my patch is fully correct. At least i did not notice any extra problems over those already known when using fglrx 9-6 with 2.6.29+.

[mahuyar]

Do you know if the problem lays within their architectural changes or new EXT4 filesystem patches? Well, I guess it is still early to judge since it is only at RC2.

[FireBurn]

Bug Reports

Out of interest did you raise a bug report for the ext4 regression?

Maybe you could write an article on how bug reports are dealt with and hopefully fixed

Would be a lot more interesting than looking at graphs from the PTS

Mike

[mendieta]

Same here. I am running 2.6.31rc2 on my eeepc (because it helps a lot with the Intel IGP). But boot time to a full KDE4 login (mostly reads from the ext4 fs on the SSD) went up from 86 secs to 105 secs or so. All this in Kubuntu 9.04.

I hope these fs regressions get worked out before the kernel goes golden!