[geeks] Re: kernel scalability....

[David Cantrell]

On Sun, Sep 09, 2001 at 01:59:03PM -0400, Greg A. Woods wrote:

>                  take the reasons as you will, but you'll likely never
> see any decent sized 390 or similar architecture machine running a linux
> kernel on the bare hardware and handling anywhere near the load

'Load' defined how?  If it's in terms of raw I/O then sure, VM or OS390
would win every time.  But it would be easy to concoct an equally
pointless benchmark which it would lose every time.

> that you can push through the same machine with multiple VM partitions.

That argument is also bogus.  You are comparing OSes designed for
completely different tasks and which target completely different
architectures*.  The fact that Linux has been ported to the 390 should
not be taken to imply that it is optimised for it.  I don't think
anyone makes that claim.

> The real reason why people run thousands of linux partitions under VM is
> because they can

Really? I thought that the few commercial installations did it because
it gave them a competitive advantage.  Are you trying to say that they
spent those millions on hardware just so they could *play* with it?

>                  and because doing so is much less expensive than any
> other option.

Ahhh, so they're actually doing it because it's the right tool for
the job.  Presumably, for those particular tasks, OS390 or VM can't
handle the load - either that or it's so much of a bitch to use that
the point is moot anyway.  I suspect a bit of both.

This second reason, I am quite sure, is far more important than the
first.

>                The physical and operational requirements necessary to
> handle the same load with linux kernels running on bare metal of any
> kind would be much larger and thus much more expensive.  I.e. that does
> definitely mean that the linux kernel cannot, alone, scale into the
> upper stratosphere of large computer systems.

Can't argue with you there, but see below ...

> Personally of course if I were to get into a situation where I would be
> buying or leasing such a solution from IBM I'd find it easy to also pay
> them the additional price necessary to port NetBSD to run under VM.
>
> Of course *BSD won't scale all that well either, at least not without a
> lot of tweaking and hacking.  :-)

The amount of tweaking and hacking would, IMO, be so large that it would
be just as practical to do the same work to the Linux kernel.

> The likes of IBM and Amdahl have spent many tens of man-years making
> Unix scale even half as well as VM on machines with that kind of
> architecture, and yet still they recommend running it under VM.

So perhaps the flaw is not with Linux, but with the fundamental concepts
underlying any Unix-a-like.  Seems pretty reasonable to me, I wouldn't
expect a mini- or micro-computer OS like a Unix to perform well when
ported to either the very high end (eg a 390) or to the very low end
(eg a Palm). And so I'm not at all surprised when both ports run into
serious problems of scalability and usability.

However, I wouldn't dream of arguing that a failure to run well in such
a Unix-hostile environment is a failure of Unix.  s/Unix/Linux/g as well.

> A more interesting comparison of linux scalability would be to compare
> it to NetBSD which runs VERY WELL on everything from the smallest of
> ARM or sh3 systems up to the largest DEC Alpha servers, not to mention
> on a lot of older hardware.

I wouldn't say it's a particularly interesting comparison to make really.
Performance and scalability of the OS are not the be-all and end-all of
deciding what to use.  Most of the bottlenecks I have come across are
in applications, not in OSes.

* - and I don't just mean 'architecture' in the sense of x86, sparc and
arm being different architectures.

-- 
David Cantrell | david at cantrell.org.uk | http://www.cantrell.org.uk/david

  This is nice.  Any idea what body-part it is?