Required Swap for recent 2.4 kernels?

Required Swap for recent 2.4 kernels?

[Quim K. Holland]

It seems that in early days of 2.4.x, the total
usable VM used to be not close to sum of the
physical RAM and swap.  Is this still true for
post 2.4.10 kernels?

Back in January 2001, in response to a question
about ``recommended swap for 2.4.x'', Rik van Riel
answered (quotes '>' are from the messagey by
alex@foogod.com):

   It has. We now leave dirty pages swapcached,
   which means that for certain workloads Linux
   2.4 eats up much more swap space than Linux
   2.2.

   On the other hand, if you almost never used
   swap under Linux 2.2, you probably won't be
   using it under 2.4 either.

   > 2) Subtract the amount of RAM you have
   >    (believe it or not, the more RAM you have,
   >    the less swap you need.  Imagine that).

   For Linux 2.4, it may be better to substract a
   bit less, because of the issue above.

   If you have a very swap-intensive workload, you
   may end up with 90% of your memory being
   "duplicated" in swap, in which case this rule
   doesn't work.

My understanding is that this answer is not
applicable to post 2.4.10 kernels, but if so then
would the total usable VM be <physical RAM> +
<swap> - <kernel itself> - <kernel overhead that
is not proportional to the VM workload>?


________________________________________________________________
Nokia 5510 looks weird sounds great. 
Go to http://uk.promotions.yahoo.com/nokia/ discover and win it! 
The competition ends 16 th of December 2001.

Re: Required Swap for recent 2.4 kernels?

[Alan Cox]

> My understanding is that this answer is not
> applicable to post 2.4.10 kernels, but if so then
> would the total usable VM be <physical RAM> +
> <swap> - <kernel itself> - <kernel overhead that
> is not proportional to the VM workload>?

Assuming it hasn't changed with the Andrea VM then the behaviour which we
switched was how the machine handles getting close to running out of swap.

The old 2.4 behaviour was that on swapping a page in from disk we always
left it in swap as well. This is a win since if the page is not dirtied
before we have to swap it out again we avoid a disk write. However it can
mean that an entire copy of RAM is sitting in swap. So a box with 1Gb of
RAM and 2Gb of swap might use 1Gb of swap and gain nothing in total
available virtual memory

The newer behaviour Rik added to the Riel VM was that when we got to about
90% of swap full we began freeing up swap pages also in RAM. That meant that
as things got tight we would migrate to a situation where data was either
in swap or ram but not both, so you went from 2Gb total usable VM to 3Gb
total usable VM on a 1Gb box with 2Gb of swap.

Alan