Re: VM patches in 2.6.4-rc1-mm2

[Mike Fedyk]

Nick Piggin wrote:
> Mike Fedyk wrote:
>
> > Andrew Morton wrote:
> >
> > > Mike Fedyk <mfedyk@xxxxxxxxxxxxx> wrote:
> > >
> > > > Most of the previous 2.6 kernels I was running on these servers 
> > > > would be lightly hitting swap by now.  This definitely looks better 
> > > > to me.
> > >
> > >
> > >
> > >
> > > It sounds worse to me.  "Lightly hitting swap" is good.  It gets rid 
> > > of stuff,
> > > freeing up physical memory.
> >
> >
> >
> > Andrew, it looks like you're right.  This[1] server doesn't seem to be 
> > hitting swap enough.  But my other[2] file server is doing great with 
> > it on the other hand (though, it hasn't swapped at all).
>
> Just curious, what makes you say [1] isn't hitting swap enough and [2]
> is OK? The graphs are better now, by the way. Thank you.

Well, with [1], I know most of those apps aren't being used, so I'm sure 
it should be hitting swap more.

And with [2], it isn't fair since I'm just happy it's not swapping 
throughout the day and all of userspace is in ram, but it's probably not 
good either.  And I'm sure there are some gettys and such that aren't 
being used, so they should have swapped out by now.

> > Maybe a little tuning is in order?
> >
> > Any patches I should try?
>
> Mainline doesn't put enough pressure on slab with highmem systems. This
> creates a lot more ZONE_NORMAL pressure and that causes swapping.

Yep, saw that.  Especially with 128MB Highmem (eg, 1G RAM)

> Now with the 2.6 VM, you don't do any mapped memory scaning at all

You mean 2.6-mm?

> while you only have a small amount of memory pressure. This means that
> truely inactive mapped pages never get reclaimed.

If I have enough pressure, they will be eventually?  But my caches will 
still be smaller than optimal, right?

> The patches you are using do not address this. My split active list
> patches should do so. Alternatively you can increase
> /proc/sys/vm/swappiness, but that isn't a complete solution, and might
> make things too swappy. It is a difficult beast to control.

Has akpm said that he would be including the active split patch in -mm?

Do you have a patch against -mm (you wrote to ask for your latest...)?

Mike
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/