Re: [rfc] SLQB: YASA

[Nick Piggin <npiggin@suse.de>]

On Thu, Apr 03, 2008 at 11:24:00AM +0300, Pekka Enberg wrote:
> Hi Nick,
> 
> On Thu, Apr 03, 2008 at 09:57:25AM +0200, Nick Piggin wrote:
> > > It's a completely different design of the core allocator algorithms
> > > really.
> > >
> > > It probably looks quite similar because I started with slub.c, but
> > > really is just the peripheral supporting code and structure. I'm never
> > > intending to try to go through the pain of incrementally changing SLUB
> > > into SLQB. If SLQB is found to be a good idea, then it could maybe get
> > > merged.
> 
> On Thu, Apr 3, 2008 at 11:13 AM, Nick Piggin <npiggin@suse.de> wrote:
> >  And also I guess I don't think Christoph would be very happy about
> >  it :) He loves higher order allocations :)
> >
> >  The high level choices are pretty clear and I simply think there might
> >  be a better way to do it. I'm not saying it *is* better because I simply
> >  don't know, and there are areas where the tradeoffs I've made means that
> >  in some situations SLQB cannot match SLUB.
> 
> So do you disagree with Christoph's statement that we should fix page
> allocator performance instead of adding queues to SLUB?

It's not just adding queues to SLUB, by any means (SLUB effectively
already has queues anyway, with it's MIN_PARTIAL thing).

I think some page allocator performance can be improved (see my patch
to remove the atomic refcounting for example). But in other cases
the page allocator just has to do a lot more work and fixing it would
just involve removing some of those things.

> I also don't
> think higher order allocations are the answer for regular boxes but I
> can see why they're useful for HPC people with huge machines.

I don't disagree with what you say. SLQB doesn't prevent them from being
used.