On Tue, 2 Jul 2002, dfox wrote:

> > No, I wouldn't say it's all that ridiculous. Of course these things are
> > relative: what you do with your machines may not require you to ever
> 
> It's not ridiculous either, unless you're thinking of emailing it
> somewhere :).

:D

> 
> > drop into swap. Kernels around 2.4.10 and earlier had serious problems
> > with memory, and requiring a few extra meg of memory would not
> 
> I'm not sure I agree, but that's a minor point. Surely some of these
> kernels had wonky use of swap, but AFAIK all linux kernels viewed swap
> as additive, rather than as 'backup store' on BSD systems. BSD is why
> most older people quote the 1.5 X as much as RAM, or even 2X swap for
> what you have in RAM rule. Of course, recent kernels have been getting
> more well behaved; I seem to remember that this was just an anomaly of
> some 2.4.x kernels, though.

I did some looking in the kernel archives and found the following:
Kernel versions from 2.3.something to 2.4.10pre9aa1 required *at 
minimum* the same amount of swap space as physical memory. This version
was when the Arcangeli patches appeared and (from his release note)
provide "swap+ram of available virtual memory".

Link: http://kt.zork.net/kernel-traffic/kt20011001_135.html#2

It does seem to have improved recently, though there are no firm
benchmarks on either side yet. Hmm...this could be an interesting little
project to stress out the vm. We'd need a little program that would grab
memory and not let it go, maybe mark each page dirty so that it stays in
the working set... I found some example code in one of Moshe Bar's
columns (Linux Kernel Pillow Talk). 

> 
> 
> My favorite swap related anecdote (true story - this actualy happened:)
> A friend of mine (who is a fairly light Linux user, had migrated from
> a Sinclair with 128k or so of RAM to a 386 with 4 megs. He thought 4
> megs was more room than he'd ever need. He didn't set up a swap partition
> at all - and then look what happened when he tried to run 2 copies of
> emacs at the same time. The kernel took *46 minutes* just to respond to
> the ctrl-x ctrl-c commands to quit one of those instances.) And of course,
> this was all in-memory thrashing since there wasn't any swapping to disk.

Pretty, isn't it? 
BTW, have you played with the overcommit settings in /proc with a
swapless system?


Want to buy your Pack or Services from MandrakeSoft? 
Go to http://www.mandrakestore.com

Reply via email to