Benchmarks Compared For Kernels 2.4.25 and 2.6.4

Josh MacLean writes "I noticed a link to an interesting article over at OSNews comparing performance of Linux kernels 2.4.25 and 2.6.4. While the workstation benchmarks are rather mundane, the server benchmarks (including Apache and MySQL) proved to be quite interesting. The rest of the Linux geeks might appreciate this. The comment thread that's linked on the last page is turning out to be relatively amusing as well."

Kernel Preempt

[MadChicken]

Interesting that they turned off the preemptible kernel, in light of what's being said on the kernel lists. I'd like to see this same suite of benchmarks with that feature turned on...

Re:Kernel Preempt

[Anonymous Coward]

IIRC, it was about 10% faster. Not all that much.

Re:Kernel Preempt

[ttfkam]

If flipping a compile switch could increase speeds in my applications by 10%, you'd better believe that I'd do it. That's what I'd call "low hanging fruit."

Re:Kernel Preempt

[MadChicken]

Ah but what was being discussed on the list was that kernel preempt makes the kernel SLOWER. I believe it was Andrea Arcangeli that was saying it was slow.

Lessee.... link, link... ah, there: link [kerneltrap.org]

Very Interesting

[TyrelHaveman]

I have wondered for some time how much of an advantage the 2.6 kernel would be. This clearly shows--at least for dual Xeon systems--that the 2.6 kernel does some excelent stuff for Samba, and seems great for MySQL [mysql.org], while only marginally better for the other tests (and sometimes worse). Someone should do similar testing with various processors. I would be interested in seeing how different it is for the Pentium 4 (with hyperthreading) and Athlon XP, especially.

Nice comparison, same results here

[mnmn]

I compiled and ran the 2.6.4 on an eSeries xServer 235 with 6 10k speed disks on RAID5, 2x HT Xeons (3.02GHz), 3 GB ram... and it crashed. The symbols didnt make much sense either.

Ive been using 2.6.3 and 2.6.4 on two other machines with great results, and running oracle on them now. I've been impressed with newer results, I thought the 2.4 ran the system close to wire speed. I've seen most of the performance difference on my SMP machine (2x Pentium3 550MHz) and filesystem (9GB SCSI 80mbps using XFS), but almost no difference on networking, video and the likes.

I'll still keep using 2.4 for full-production systems, well, partly because UML is not mature (or even available) on 2.6 yet.

Re:I installed 2.6

[Vlad_the_Inhaler]

Well, I put some new hardware in (USB 2.0 PCI Card) and now 2.4 no longer runs correctly, something in SuSE's startup procedure locks up.

2.6.4 runs fine though.

statistics

[Hatta]

One second isn't exactly a huge gap, but this was averaged over five runs, so I will have to give the nod to the venerable 2.4 kernel here.

Well at least he knows how to take a mean. Why is it that benchmarks never include any error? Is it so hard to put errorbars on a chart? Is a simple T test too much to ask?

Apache 1.3?

[ttfkam]

Apache 2.0 has a worker module that uses both threads and processes on Linux. 1.3 just preforks. Since one of the big improvements in 2.6 was threading support and since Apache 2.0 can run in either mode anyway, why test 1.3?

If you want to test something more "stable," why bother testing the 2.6 kernel? After all, the 2.6 kernel is newer than Apache 2.0.

Operating System: Gentoo Linux 1.4.3.13

[Spunk]

What?

I run Gentoo and I wasn't aware that there were version numbers... :) Where's he getting that number?

Re:Operating System: Gentoo Linux 1.4.3.13

[Anonymous Coward]

/etc/gentoo-release?

Re:Operating System: Gentoo Linux 1.4.3.13

[Spunk]

Huh. Interesting. Then I have 1.4.3.13 too :)

The timestamp matches my last emerge of sys-apps/baselayout.

Re:Operating System: Gentoo Linux 1.4.3.13

[MarcQuadra]

Probably means Gentoo-1.4 built March 13th. Still, bad naming, I mark my gentoo stage builds and portage snapshots like so:

stage1-20040309.tar.bz2
stage1-20040313.tar.bz2
stage1-20040313-mpd1.tar.bz2
etc.

using ISO-format datestamps like that in filenames is a godsend for anyone who does multiple builds of stuff, they line up nicely no matter what OS you're using to browse the files.