Re: The Kernel (fwd)

[jfm2@club-internet.fr]

> 
> On 25 Jan 1998 jfm2@club-internet.fr wrote:
> 
> > 1) Bogomips have nothing to do with our problem: we are speaking of
> > memory usage.  That is related with _SWAPPING_.  By reducing kernel
> > size you will make your machine faster when short of memory because
> > you will have lesser swapping.
> 
> I doubt it ... how much swapping do think is going on anyway? Besides, if
> we take out 200K or even 500k (I seriously doubt it would be that high but
> for the sake of the discussion ... ok).  Look at your system ... how much
> do you have swapped out now.  What processes are they?  Chances are those
> are getty processes on unused VT's that have been swapped out.  The fact
> is that on a system of >32MB, there is just not a whole lot of swapping
> going on and once you start a really big program running, that 500k is not
> going to make any difference because you are probably short more than that
> anyhow.
> 
> Look at it this way.  Removing those drivers gives exactly the same
> performance boost of adding 2 256k simms. (Provided that you can get 500k
> out of there but I doubt it).  How much performance gain do you get from
> adding 1/2Meg of RAM?  If you have 4MEG, you probably get a lot.  If you
> have 32MEG, you will probably never notice it.
> 

You see Linux is supposed to be a system lean and fast.  With time it
has more and more drivers.  The SCSI drivers in 2.0 are over 500K: I
KNOW it by lloking at what I get in the modules directory.  In 2.2
there will be far more.  How are we going to call Linux when in 2.4 or
3.0 a kernel with compiled in drivers will be 32 Megs?  Slowinux?

Second: You ever point at a new machine with 32 Megs.  At work I have
only 16 Megs and because I am not willing to pay memory from my pocket
it is my boss who decides how much memory I have.  And the cheap
memory argument is valid on Pentium motherboard but 8 bit SIMMS the
only usable on old boxes was _VERY_ expensive last time I checked.

Or are people supposed to change computers every year?  In that case
mail me I will buy some Intel shares.

> That is what I am getting at.  If you can document a significant
> performance boost, that would be worth considering. I am just skeptical of
> ACTUAL performance boost gained from this as opposed to problems.
> 

When you have 6 Megs left after your kernel is loaded and the working
set of your programs (Read books to know what is a working set) is 6
Megs 500 you will begin to thrash.  Not swap, thrash: when you spend
99% of your time paging.

> Have you asked the Debian kernel maintainer why those were left in there
> ... nevermind, I will do it.  For all I know, they MIGHT have tried this
> but run into problems. 
> 

And why don't you ask to people from the distributions who are using
it?  Do you think than the Debian mantainer is automatically better
than them? 

> There is a THEORETICAL performance boost IF you are constantly thrashing
> things in and out of swap that are smaller than 500k in size. Once you
> start swapping more than this in and out of RAM (A 4MB news history file
> for example) that 1/2MEG is meaningless.
> 

I, one of the dev-install leaders, who has been investigating fully
modulars kernels since around 1.3.50 because they were fundamental for
carrying Linux to end users , think than this falls in the install
area not in yours.  And also than I know enough to be able to think by
myself without taking the word of any mantainer as sacred wisdom.


> > 2) Adding RAM or reducing the kernel size will have ZERO effect if all
> > your processes fitted in memory.
> 
> Correct, so until you get to the point of swapping, there is no
> performance gain.  You DO gain performance in the range between the time
> you start swapping up until you reach a point where you are swapping out
> >500k. At that point performance is again identical to the non-stripped
> kernel.
> 

Yeap.  If you are in the easy case of having enough memory for the
process subject to the condition of swapping sleeping processes.

No.  If the active parts of the non-sleeping processes are bigger than
your memory.  Then 500K will make a big difference.  And if you are
lucky enough the 500K will allow them to fit entirely into memory so
you will not have thrashing.



> 8MB SIMMS are $12 here.  For $24 you get a 32 times performance boost than
> stripping the SCSI drivers out of the kernel.
> 

Expensive memory in 386, 486.  Full motherboards.  What when you use
Linux at work?

I have no SIMM manufacturer shares and I know what is to use 8 M
boxes.  I am not going to deprive from 500K a user who cannot
recompile his kernel just by incompetency.

-- 
			Jean Francois Martinez

==================== The Linux.  Use the Linux, Luke! =======================