[Beowulf] Racing kernels

Mark Hahn hahn at physics.mcmaster.ca
Mon Apr 4 23:09:21 PDT 2005

> By "racing" I mean removing all unnecessary components (SCSI, USB et 
> al), tweaked buffer values, pre-empting, stripped, helium filled, 
> brushed aluminum etc?

sure.  why build things you're not going to use?  and why bother 
with modules if you're going to need that driver for your root
filesystem anyway?  of course, this does presume that you have 
some reason to change kernels (I track kernel.org - updating 
whenever it looks like something interesting has been added.)

but this is mostly for ease of installation, rather than racing.

> Yes, it's a loaded question. I'm contemplating performance tweaking when 

I've never noticed that custom kernels make any real difference
in performance.  even parameter tweaking is often surprisingly useless.

> I'd like to think we're giving priority to stability. But still... I 
> think there's scope for "safe" improvement?

I don't think minimal, monolithic kernels are noticably faster,
but they ARE obviously just as safe.  even if you are religiously 
committed to a kernel annointed by RH, you can safely recompile
the kernel.  backing out of, oh, say, the 4k-stack patch is hardly
going to screw up your disk driver.

kernel modules are way cool if you're a driver hacker or you 
really do have peripherals that you only use occasionally.
peripherals with really big drivers, I suppose.  I've been tempted
to compile in the myrinet driver, for instance, since I never 
want to unload it, and it ties itself to the kernel version anyway.

> Also need to include the type of target app(s) I s'pose. My Nbody apps 
> are smokin' math stuff but the IO is quite light. So there'd be a fair 

you might try turning HZ down to 100.  if it makes a measureable difference,
please to post your results...

regards, mark hahn.

More information about the Beowulf mailing list