Message-ID: <20070806225718.GA16209@skynet.ie>
Date:	Mon, 6 Aug 2007 23:57:18 +0100
From:	mel@...net.ie (Mel Gorman)
To:	Christoph Lameter <clameter@....com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>, Andi Kleen <ak@...e.de>,
	Lee.Schermerhorn@...com, kamezawa.hiroyu@...fujitsu.com,
	linux-mm@...ck.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] Apply memory policies to top two highest zones when highest zone is ZONE_MOVABLE

On (06/08/07 15:31), Christoph Lameter didst pronounce:
> On Mon, 6 Aug 2007, Mel Gorman wrote:
> 
> > > So where do we stand on this?  We made a mess of NUMA policies, and merging
> > > "grouping pages by mobility" would fix that mess, only we're not sure that
> > > we want to merge those and it's too late for 2.6.23 anwyay?
> > > 
> > 
> > Grouping pages by mobility would still apply polciies only to
> > ZONE_MOVABLE when it is configured. What grouping pages by mobility
> > would relieve is much of the motivation to configure ZONE_MOVABLE at all
> > for hugepages. The zone has such attributes as being useful to
> 
> Ultimately ZONE_MOVABLE can be removed. AFAIK ZONE_MOVABLE is a temporary 
> stepping stone to address concerns of about defrag reliability. Somehow 
> the stepping stone got into .23 without the real thing.
> 
> An additional issue with the current ZONE_MOVABLE in .23 is that the 
> tentative association of ZONE_MOVABLE with HIGHMEM also makes use of large 
> pages by SLUB not possible.
> 

Pretty much. The use of ZONE_MOVABLE really only applies to hugepages
and potentially memory hot-remove right now.

> > There are patches in the works that change zonelists from having multiple
> > zonelists to only having only one zonelist per node that is filtered based
> > on the allocation flags. The place this filtering happens is the same as what
> > the "hack" is currently doing. The cost of filtering should be offset by the
> > reduced size of the node structure and tests with kernbench, hackbench and
> > tbench seem to confirm that. This will bring the hack into being line with
> > what we wanted with policies in the first place because things like MPOL_BIND
> > will try nodes in node-local order instead of node-numeric order as it does
> > currently.
> 
> I'd like to see that patch.
>  

I'll find the time to get it implemented this week. I've been
prioritising anything that looked like a bug recently so it languished
on the TODO pile.

> > >From there, we can eliminate policy_zone altogether by applying policies
> > to all zones but forcing a situation where MPOL_BIND will always contain
> > one node that GFP_KERNEL allocations can be satisified from. For example,
> > if I have a NUMAQ that only has ZONE_NORMAL on node 0 and a user tries to
> > bind to nodes 2+3, they will really bind to nodes 0,2,3 so that GFP_KERNEL
> > allocations on that process will not return NULL. Alternatively, we could
> > have mbind return a failure if it doesn't include a node that can satisfy
> > GFP_KERNEL allocations. Either of these options seem more sensible than
> > sometimes applying policies and other times not applying them.
> 
> We would still need to check on which nodes which zones area available. 
> Zones that are not available on all zones would need to be exempt from 
> policies. Maybe one could define an upper boundary of zones that are 
> policed? On NUMAQ zones up to ZONE_NORMAL would be under policy. On x86_64 
> this may only include ZONE_DMA. A similar thing would occur on ia64 with 
> the 4G DMA zone. Maybe policy_zone could become configurable?
>  

A sensible upper-boundary would be if GFP_KERNEL can be used on zones within
that that node or not. gfp_zone(GFP_KERNEL) provides that sort of information
and would resolve to ZONE_NORMAL on NUMAQ, ZONE_DMA on ia64 etc. Prehaps
that would not work out for GFP_DMA32, I'm not 100% sure at the moment
(it's late and this is meant to be a holiday, sue me :) )

This discussion is independent of one-zonelist-per-node which is the
stepping stone between where we are now and getting rid of policy_zone
altogether.

> > I'm for merging the hack for 2.6.23 and having one-zonelist-per-node
> > ready for 2.6.24. If there is much fear that the hack will persist for too
> 
> Why not for .23? It does not seem to be too much code?
> 

I'm working under the assumption that if it's not a bug-fix, you can't
get it in after the merge window closes. I've seen complaints before where
"bug-fixes" were adding features which one-zonelist-per-node may be preceived
by some people to be. Perhaps the rules will flex for this patch when it comes
out, perhaps not. I made the assumption that the least invasive bug-fix was
sensible outside of the merge window and that's what this hack is.

> > long, I'm ok with applying policies only to ZONE_MOVABLE when kernelcore=
> > is specified on the command line as one-zonelist-per-node can fix the same
> > problem. Ultimately if we agree on patches to eliminate policy_zone altogether,
> > the problem becomes moot as it no longer exists.
> 
> We cannot have a kernel release with broken mempolicy. We either need the 
> patch here or the one-zonelist patch for .23.

I'll get a sensible version of one-zonelist ASAP. Prehaps we'll end up
just going with that altogether if no performance issues are evident in
testing.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists