Linked by Eugenia Loli on Sun 21st Sep 2003 18:16 UTC, submitted by Jeremy Andrews
Linux Anyone interested in squeezing maximum performance out of their GNU/Linux operating system and willing to compile a new kernel will be interested in this KernelTrap article about Andrew Morton's -mm patchset. The patchset currently offers better stability and performance than the mainline 2.6.0-test kernel, as well as containing numerous functionality enhancements. Much of the additional functionality is described, as well as providing simple step by step installation instructions. Reading this, one can understand why Linux creator Linus Torvalds has chosen Andrew to soon become the 2.6 maintainer.
Order by: Score:
v Solution to linux kernel performance problems:
by yerma on Sun 21st Sep 2003 18:43 UTC
Good article.
by Greg on Sun 21st Sep 2003 18:57 UTC

Does anybody know how the latest -mm compares to Robert Love's -love patchset?

v kernel 2.6
by Arkady on Sun 21st Sep 2003 19:11 UTC
My quick hdparm test
by Anonymous on Sun 21st Sep 2003 19:25 UTC

I'm using test5-mm3 and here are my results, however these are for test5-mm1, since I'm not in Linux right now I can't do another test, they are pretty impressive to say the least, and I think most of you will think so too. This is on an ordinary IBM 85gig ATA 100 harddrive.

Timing buffer-cache reads: 1004 MB in 2.01 seconds = 500.33 MB/sec
Timing buffered disk reads: 136 MB in 3.01 = 45.13 MB/sec

I use to get in the low 4s or even 3s on a 2.4 kernel

@Greg : Love Sources
by Lovechild on Sun 21st Sep 2003 19:28 UTC

Love sources has nothing to do with Robert Love.

It's my personal little project which started to test Nick Piggins Scheduler Policy patch as well as the newest CFQ patch, but now I seem to have gotten a lot of patches included.... (Psst. Nick hurry up with that Nick v16 release - my users are starting to look.... hungry ;) )

One of my users used Contest to test -love against -mm and vanilla and in most tests -love seems to do as good or better as -mm, and vanilla is completely blown away by both patchsets.

You can read this thread:
http://forums.gentoo.org/viewtopic.php?t=81758 for more information on the patchset.

But please remember I'm by no standards a kernelhacker - I merely collect patches and apply them to the milestone releases of linus' vanilla kernel. It's work a slightly trained monkey can handle, and the only reason they are in circulation is because I think it's fun.

RE: My quick hdparm test
by Anonymous on Sun 21st Sep 2003 19:33 UTC

I went to into Linux to see the improvement between mm1 and mm3 and here are the new results using mm3

Timing buffer-cache reads: 1036 MB in 2.01 seconds = 516.27 MB/sec
Timing buffered disk reads: 138 MB in 3.01 = 45.45 MB/sec

Not a huge improvement but a little higher then last time.

RE: @Greg : Love Sources
by Greg on Sun 21st Sep 2003 19:39 UTC

Oh! Sorry about that!

Anyway, I'm just trying yours out this weekend, and it performs nicely--it's good to be able to try Reiser4, though I haven't gotten around to it.

Good job!

My experience with kernel patches
by ProgrammmerMan on Sun 21st Sep 2003 19:48 UTC

I've tested the 2.6.0-test5 kernel with Con Kolivas' interactivity patch. I've also tested Con's 2.4.21 and 2.4.22 patches and so far nothing can match the gentoo-sources-2.4.20-r5-aavm patch for speed. (I'm using Mandrake 9.1 and GCC 3.3.1, BTW.) Just watching the init scripts running at bootup and how long XFree86 and Mozilla Firebird take to launch, there's no comparison. Considering that this patch uses Con's performance enhancements, I'm not too confident that it can do any better. But maybe I'll test it out and see what it can do.

If anyone is interested in a mini-howto on using the gentoo-sources patch, speak up and I'll try to whip something up.

RE:
by Hemal on Sun 21st Sep 2003 19:57 UTC


Buffered timings are not very accurate and can fluctuate greatly. It would be better to use unbuffered timings to compare results with other kernels.

>I went to into Linux to see the improvement between mm1 >and mm3 and here are the new results using mm3
>
>Timing buffer-cache reads: 1036 MB in 2.01 seconds = >516.27 MB/sec
>Timing buffered disk reads: 138 MB in 3.01 = 45.45 MB/sec
>
>Not a huge improvement but a little higher then last time.

bk tree
by Jon Smirl on Sun 21st Sep 2003 22:05 UTC

Is there a bk tree for mm, or are there only patches?

2.6.0-test-mm tree for newer mobile devices
by butters on Sun 21st Sep 2003 22:47 UTC

If you are an owner of a P4M-based (Centrino) laptop, especially the ones with integrated i855GM graphics, I highly recommend these kernels (or any 2.6.x kernels, but mm is definitely the way to go). They have kernel DRI support for i855GM and support for ACPI and CPU frequency throttling for Intel Enhanced Speedstep. The pcmcia-cs package might throw some configure errors if you compile it from source under a 2.6.x kernel with the new modutils replacement, but installing the older modutils or compiling under your old 2.4.x kernel will fix this minor issue.

mm3 rocks
by hirisov on Mon 22nd Sep 2003 00:20 UTC

I've tested some kernel's w/ the contest package. These were 2.4.22, 2.4.22-ac3, 2.6.0-test5, 2.6.0-test5-mm3, 2.4.21-wolk9s and 2.4.18-bf24 (as it's in the debian stable). I won't paste here long tables, but believe me mm3 rocks and was the best among them in this test. 2.4.22-ac3 and 2.6.0-test5 can keep up w/ it in most tests, but under heavy i/o load the new vm and vfs simply rocks, the 2.6 kernels blows away every older one. mm3 is faster than 2.6.0-test5 vanilla in all the tests.



SO
by Beavis on Mon 22nd Sep 2003 01:25 UTC

Will this eventually merge into the main kernel?

re: so
by debman on Mon 22nd Sep 2003 02:59 UTC

yeah, in the 2.7 branch maybe....2.5/2.6 is closed to merges of new code.

Re: debman
by Wade on Mon 22nd Sep 2003 08:37 UTC

It's not closed to merges, just closed to huge changes. They are still accepting a lot of bugfixes/sorely-needed-features. The -mm patches will make it into 2.6 once Andrew is the maintainer.