VM working much better in 2.4.8 than before

VM working much better in 2.4.8 than before

[misty-]

	I've said it before, I'll say it again - you guys deserve a
'this is a great thing' report once in a while. This is such a report :)
>From 2.4.7 to 2.4.8 the greatest effect I've seen so far is that my swap
use is quite minimal - This affects my 486 quite noticably, I don't have
exact figures but I'm guessing it's startup time is 1/3 of normal.
Shocked the heck out of me, and it's using only 10MB of swap when
idle where it used to use at least 15-20MB. Also I've noticed much less
swap activity during heavy use, which is really helping both my old 486
in a extremely noticable way (hard disk is a huge bottleneck on that
poor demented trashcan :) and my K6-III which has a udma 66 disk... When
it doesn't have to pay attention to writing to the swap partition, it
obviously can be doing other things. Which helps. Really! :)

Thanks for the improvements, all of you,

Tim McGrath

(formerly at tcm@nac.net)

Re: VM working much better in 2.4.8 than before

[Helge Hafting]

misty-@charter.net wrote:
> 
>         I've said it before, I'll say it again - you guys deserve a
> 'this is a great thing' report once in a while. This is such a report :)

2.4.8 is good indeed.  My vmstat window tells me it ties up
about 40M less in page cache than earlier kernels, possibly an
effect of use-once.  (This is a 128M desktop machine.)

updatedb run still manages to use some swap, but the machine
is no longer sluggish in the morning. :-)

Helge Hafting

Re: VM working much better in 2.4.8 than before

[Daniel Phillips]

On Monday 13 August 2001 10:41, Helge Hafting wrote:
> misty-@charter.net wrote:
> > 
> >         I've said it before, I'll say it again - you guys deserve a
> > 'this is a great thing' report once in a while. This is such a report :)
> 
> 2.4.8 is good indeed.  My vmstat window tells me it ties up
> about 40M less in page cache than earlier kernels, possibly an
> effect of use-once.  (This is a 128M desktop machine.)
> 
> updatedb run still manages to use some swap, but the machine
> is no longer sluggish in the morning. :-)

Yes, those would be the expected effects of use-once, in fact it was
"morning after updatedb" question that got me started on it.

BTW, keep that 486, it's a great reality check!

--
Daniel

Re: VM working much better in 2.4.8 than before

[Alan Cox]

> Yes, those would be the expected effects of use-once, in fact it was
> "morning after updatedb" question that got me started on it.

updatedb is also absolutely fine if you just work with the existing VM
and up the inode pressure a little. I'm still very unconvinced by use-once.

Alan

Re: VM working much better in 2.4.8 than before

[Rik van Riel]

On Mon, 13 Aug 2001, Alan Cox wrote:

> > Yes, those would be the expected effects of use-once, in fact it was
> > "morning after updatedb" question that got me started on it.
>
> updatedb is also absolutely fine if you just work with the existing VM
> and up the inode pressure a little. I'm still very unconvinced by
> use-once.

Use-once has a number of theoretical disadvantages too:

1) newly read in pages are evicted earlier, this means
   readahead pages will either evict each other or the
   amount of readahead done might need to be shrunk
   -- the current readahead code is not prepared for this,
      use-once could lead to more disk seeks being done

2) since we add new pages to the inactive list, VM
   balancing is faced with a really strange situation ;)

Yes, these things are solvable, but not without redesigning
major parts of the VM balancing to do things which have never
been done before. I'm not sure 2.4 is the time to do that.

regards,

Rik
--
IA64: a worthy successor to i860.

http://www.surriel.com/		http://distro.conectiva.com/

Send all your spam to aardvark@nl.linux.org (spam digging piggy)

Re: VM working much better in 2.4.8 than before

[Daniel Phillips]

On August 13, 2001 05:52 pm, Rik van Riel wrote:
> On Mon, 13 Aug 2001, Alan Cox wrote:
> > > Yes, those would be the expected effects of use-once, in fact it was
> > > "morning after updatedb" question that got me started on it.
> >
> > updatedb is also absolutely fine if you just work with the existing VM
> > and up the inode pressure a little. I'm still very unconvinced by
> > use-once.
> 
> Use-once has a number of theoretical disadvantages too:
> 
> 1) newly read in pages are evicted earlier, this means
>    readahead pages will either evict each other or the
>    amount of readahead done might need to be shrunk
>    -- the current readahead code is not prepared for this,
>       use-once could lead to more disk seeks being done

Have you actually seen this happening?
 
> 2) since we add new pages to the inactive list, VM
>    balancing is faced with a really strange situation ;)
> 
> Yes, these things are solvable, but not without redesigning
> major parts of the VM balancing to do things which have never
> been done before. I'm not sure 2.4 is the time to do that.

--
Daniel