Recompiled Kernel >> Performance tests
 

Hi :)

I've recompiled my kernel a few times on differnet distro's recently to try and optimise performance.

Anyway it all went great:

Initial standard kernel sizes were ~ 1.1MB on Slackware and Red Hat, and I managed to decrease the kernel to < 800KB :D

Anyway I thought I'd try some sort of test to see if this decreased kernel size would lead to faster computation.

The results I got were a little dissapointing :(

Each test was performed in exactly the same way

*Rebooted
*Logged into Gnome
*Opened Terminal
*Switched User to root
*openssl speed

Nothing was done differently each time the process was performed

Basically the tests were done MARGINALLY quicker with the stock kernels.

I can't understand why this was the case :confused:

Is this test dependant upon the size of the kernel?

My thinking was that it must be since everything is done via the kernel.

If anyone has any explanations or advice then it would be gratefully received.

BTW - I used make oldconfig to keep the existing options the same and just removed a lot of the excess things that I don't require in the kernel such as IRDA, ISDN etc ... and also to change my CPU to the correct type - AMD

Thanks

:)

Is this test dependant upon the size of the kernel?

No, you cannot test an optimized kernel that way, and compiling a kernel so it is smaller in size does not optimize stuff a lot. Stuff that could make a kernel "run faster" could be: using different optimization flags for kernel compilation like -O2 to -O3 (warning: get to know what gcc optimization does!) or/and selecting a more "aggressive" Virtual Memory manager: see Andrea Angeli's patches, selecting a another/a more "aggressive" scheduler: Rik van Riel (IIRC), Ingo Molnar, using Dave Morton's latency/preempt patches or/and for instance trying different values for the VM sysctl subsystem tuning HD with hdparm. Much of the patch stuff is rolled up in combined kernelpatches like Con Koliva or WOLK releases btw.

You should try to find out (a little bit) more about how parts of the kernel function, like the difference between "userland" processes and syscalls, how a scheduler works and the impact of caching and VM management. To test it you need to get data that shows how everything functions while you perform tasks you know will "stress" the system with "stress", "bonnie" etc etc.

To get to see in which syscalls the kernel spends most of its time, add "profile=2" to the LILO append line, run "kerneltop" and output the top 20 at regular intervals. To get data about process status, VM, IRQ, network, disks usage etc, etc I prefer using "atsar".

Must reads:
http://kerneltrap.org
http://kernelnewbies.org/faq/
http://linuxperf.nl.linux.org/general/kerneltuning.html
http://www.nks.net/linux-vm.html
http://home.earthlink.net/~rwhron/kernel/2.4.html

Only if your fix is hardcore kernel pr0n:
http://www.kernelhacking.org/docs/ke...WTO/index.html
http://www.kernelnewbies.org/documen...ing-guide.html
http://iven.home.cern.ch/iven/linux_...ash/index.html
http://oss.hitachi.co.jp/sdl/english/lkst.html

*For any names or phrases not linked here: please Google/Freshmeat for 'em.

Thanks unSpawn

Looks like I've got some reading to do then :)

Thanks, unSpawn.
This is a GREAT post. very helpful to me. have been looking for this for a long time.