Open Discussion, kernel in production environment

Open Discussion, kernel in production environment

[Marcin Hlybin]

Hello,

I always configure and compile a kernel throwing out all unusable options and 
I never use modules in production environment (especially for the router). 
But my superior has got the other opinion - he claims that distribution 
kernel is quite good and in these days optimization has no sense because of 
powerful hadrware. 
What do you think? I have few arguments for this discussion but I wonder what 
you say. Please, try to substantiate your opinions.

And the second question: which kernel branch/version *and why* do you use in 
production environment?

Regards

-- 
 Marcin Hlybin, marcin.hlybin@swmind.com
 Sys/Net Administrator, tel. +48 12 2523 402

Re: Open Discussion, kernel in production environment

[Bernd Eckenfels]

Marcin Hlybin <marcin.hlybin@swmind.com> wrote:
> I always configure and compile a kernel throwing out all unusable options and 
> I never use modules in production environment (especially for the router). 
> But my superior has got the other opinion - he claims that distribution 
> kernel is quite good and in these days optimization has no sense because of 
> powerful hadrware. 

I think it is most often wasted time to optimize your kernels that way,
especially if you consider the ongoing work you have to put into maintenance
of it.

So unlike you plan to roll out hundreds of systems with the same image
(containing a hand made kernel) I would not suggest to do it.

Of course I am not talking about embedded systems here. For larger Servers
(NUMA, Terrayte Ram stuff like that) I also would think you might want to go
with the Vendor kernel, for better support.

For production use with things like Oracle or SAP, I would suggest to not
use uncertified kernels anyway.

Gruss
Bernd

Re: Open Discussion, kernel in production environment

[Andi Kleen]

Marcin Hlybin <marcin.hlybin@swmind.com> writes:

> Hello,
> 
> I always configure and compile a kernel throwing out all unusable options and 
> I never use modules in production environment (especially for the router). 
> But my superior has got the other opinion - he claims that distribution 
> kernel is quite good and in these days optimization has no sense because of 
> powerful hadrware. 
> What do you think? I have few arguments for this discussion but I wonder what 
> you say. Please, try to substantiate your opinions.

Try to benchmark the difference. If you can't it's wasted work.
Most likely you can't.

-Andi

Re: Open Discussion, kernel in production environment

[Arjan van de Ven]

On Mon, 2006-05-01 at 23:57 +0200, Marcin Hlybin wrote:
> Hello,
> 
> I always configure and compile a kernel throwing out all unusable options and 
> I never use modules in production environment (especially for the router). 
> But my superior has got the other opinion - he claims that distribution 
> kernel is quite good and in these days optimization has no sense because of 
> powerful hadrware. 


he's basically right; the gain you get by disabling the generic things a
distribution enabled is way down in the noise. There are some
compromises a distribution makes to keep the number of kernels they ship
down, and I suspect the worst possible case would add up to say 5%.

(on a Fedora / RHEL that would probably be a SMP system with less than
1Gb of ram)

On the plus side you get the maintenance, building and integration done
for you, including the security fixes. 

There is a third "advantage" in using a distro kernel; there is less
chance of a mistake in the sense of picking a config option that turns
out to be really bad in hindsight. 

Now that doesn't mean that I want to discourage people from building
their own kernel, far from that, but at the same time, the advantages
you get shouldn't be overstated and in a business environment it is
probably not a good use of time nowadays.

Re: Open Discussion, kernel in production environment

[Jan Engelhardt]

>> I always configure and compile a kernel throwing out all unusable options and 
>> I never use modules in production environment (especially for the router). 
>> But my superior has got the other opinion - he claims that distribution 
>> kernel is quite good and in these days optimization has no sense because of 
>> powerful hadrware. 
>
>On the plus side you get the maintenance, building and integration done
>for you, including the security fixes. 
>
>There is a third "advantage" in using a distro kernel; there is less
>chance of a mistake in the sense of picking a config option that turns
>out to be really bad in hindsight. 
>
At best pick the distro kernel .src.rpm (or equivalent). If required, you 
can still roll your own _and_ have the security fixes etc.


Jan Engelhardt
-- 

Re: Open Discussion, kernel in production environment

[David Lang]

On Mon, 1 May 2006, Marcin Hlybin wrote:

> Hello,
>
> I always configure and compile a kernel throwing out all unusable options and
> I never use modules in production environment (especially for the router).
> But my superior has got the other opinion - he claims that distribution
> kernel is quite good and in these days optimization has no sense because of
> powerful hadrware.
> What do you think? I have few arguments for this discussion but I wonder what
> you say. Please, try to substantiate your opinions.

At one point in the past I did some testing of stock 386 kernels (standard 
at the time) vs K7 optimized kernels that were stripped down and saw a 
substantial difference (30% IIRC). nowdays the distros are optimizing for 
better chips so it makes less of a difference (and there's not a lot of 
variation yet among the amd64 options)

however, by rolling your own you can avoid including features that you 
don't need, and by eliminating those features you also eliminate any 
bugs/vunerabilities that those features include (the trade-off is that 
sometimes there are dependancies that require non-obvious features and/or 
features are buggy and break things when turned off).

I got burned by distro kernels in the past, and the distro was not much 
help (not everything in the system was on that distro's 'approved 
hardware' list, and even the stuff that was hadn't been testing in that 
particular combination). this makes me cynical about the testing that the 
distro does (they can't possibly test every combination), so if I have to 
test things myself why wait the extra time for the disto to do it's 
release?

things are getting better with the distros staying close to the vanilla 
kernel, but my memory is long enough to not really trust them yet ;-)

David Lang


-- 
There are two ways of constructing a software design. One way is to make it so simple that there are obviously no deficiencies. And the other way is to make it so complicated that there are no obvious deficiencies.
  -- C.A.R. Hoare

Re: Open Discussion, kernel in production environment

[Grant Coady]

On Tue, 2 May 2006 18:06:27 -0700 (PDT), David Lang <dlang@digitalinsight.com> wrote:

>things are getting better with the distros staying close to the vanilla 
>kernel, but my memory is long enough to not really trust them yet ;-)

But the distro kernel is only there to boot the machine in order to 
compile a custom kernel?  And for reboot when one's shiny new custom 
kernel falls over :o)  

Grant.

Re: Open Discussion, kernel in production environment

[Jan Engelhardt]

>>things are getting better with the distros staying close to the vanilla 
>>kernel, but my memory is long enough to not really trust them yet ;-)
>
>But the distro kernel is only there to boot the machine in order to 
>compile a custom kernel?  And for reboot when one's shiny new custom 
>kernel falls over :o)  
>
The best systems have that 'distro kernel' also replaced,
by a self-made 'working' one :>


Jan Engelhardt
--