Subject: Re: Paging in from swap
To: Jason R Thorpe <email@example.com>
From: Luke Mewburn <firstname.lastname@example.org>
Date: 01/05/2003 11:15:14
On Sat, Jan 04, 2003 at 03:08:06PM -0800, Jason Thorpe wrote:
| On Sat, Jan 04, 2003 at 11:01:31PM +0100, Reinier Jonker wrote:
| > This does not seem logical to me. When the pages in swap are required
| > again, there is a huge performance penalty when paging them back in.
| > Wouldn't it be better to page swap pages back in during idle time if
| > the amount of free memory permits that? I do see one disadvantage, it
| > may be required to page them back out before the pages are used again,
| > but it may be possible to reduce that disadvantage with a timeout
| The reason for leaving them out in swap until they're needed again is
| because the pages were "idle" in the first place (that's how they were
| selected to go out to swap). This means that they're likely to continue
| to be "idle", and so leaving them there until they are actually needed
| saves pointless I/O.
I can see one situation where this would be useful; a laptop where
you've set the disk to spin-down on idle. On the occasions where you
fire up a big program that causes a little bit of swap, the disk spins
up and runs for a while. What would be nice (*) is if the system
would page in those pages if there's gobs of free RAM again whilst
the disk is spinning would be cool.
Of course, tuning my laptop's vm footprint, and/or adding RAM can
help this problem.
Having other features to be more aware of a disk's "idle" state would
be nice too; I've got my laptop running very quietly most of the time,
because I run noatime,nodevmtime, and I've tweaked services which
normally access the disk too often (syslogd - goes to a loghost,
postfix - with its habit of spinning up the disk every minute - stopped,
(*) yes, this is a corner case, and I'm asking for "pie in the sky"