On Mon, 31 Dec 2012, Trent W. Buck wrote:

> Anthony Hogan <[email protected]>
> writes:
>
> > Hrrmm.. interesting RE swap. My reading in the past regarding swap
> > seems to be not to starve the machine on it or stuff just dies rather
> > than slows down.. it's something to keep in mind though.
>
> IME as at 2.6, the last time I ran with swap, "slows down" means "slows
> down to the point where you can't get a shell to spawn pkill" such that
> you end up having to hard reset the entire machine.  Compared to that,
> letting the OOM killer take a guess (and probably kill off the browser
> that's causing swap thrash) seems like a good bet.  YMMV &c.
>
> Oh, for servers I do generally still give them swap, but since they are
> massively overspecced with RAM, I've never seen them actually page
> anything out.

Since SSDs can almost be written to all day every day for about 5 years
without developing dodgy cells these days, I decided to enable swap on my
SSD based machine.  And was surprised to still only get about 3MB/S swap
rates (15MB/s if I also enable zram).  There is something really fscking
screwy in Linux's swap implementation if the best it can do when trying to
swap a 1GB browser image back in from disk is 3MB/s (@ about 300 iops) off
a device that can do 80,000 iops and 300GB/s read.

It's been like that since at least the 2.4 days.  I've only *ever* gotten
3MB/s out of the si and so columns in vmstat. (at least SSD allows me to
keep using the machine for other tasks when it's busy swapping in opera).

-- 
Tim Connors
_______________________________________________
luv-main mailing list
[email protected]
http://lists.luv.asn.au/listinfo/luv-main

Reply via email to