swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Al Boldi]

People wrote:
> >> I believe the users who say their apps really do get paged back in
> >> though, so suspect that's not the case.
> >
> > Stopping the bush-circumference beating, I do not. -ck (and gentoo) have
> > this massive Calimero thing going among their users where people are
> > much less interested in technology than in how the nasty big kernel
> > meanies are keeping them down (*).
>
> I think the problem is elsewhere. Users don't say: "My apps get paged
> back in." They say: "My system is more responsive". They really don't
> care *why* the reaction to a mouse click that takes three seconds with
> a mainline kernel is instantaneous with -ck. Nasty big kernel meanies,
> OTOH, want to understand *why* a patch helps in order to decide whether
> it is really a good idea to merge it. So you've got a bunch of patches
> (aka -ck) which visibly improve the overall responsiveness of a desktop
> system, but apparently no one can conclusively explain why or how they
> achieve that, and therefore they cannot be merged into mainline.
>
> I don't have a solution to that dilemma either.

IMHO, what everybody agrees on, is that swap-prefetch has a positive effect 
in some cases, and nobody can prove an adverse effect (excluding power 
consumption).  The reason for this positive effect is also crystal clear:  
It prefetches from swap on idle into free memory, ie: it doesn't force 
anybody out, and they are the first to be dropped without further swap-out, 
which sounds really smart.

Conclusion:  Either prove swap-prefetch is broken, or get this merged quick.


Thanks!

--
Al

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Arjan van de Ven]

> IMHO, what everybody agrees on, is that swap-prefetch has a positive effect 
> in some cases, and nobody can prove an adverse effect (excluding power 
> consumption).  The reason for this positive effect is also crystal clear:  
> It prefetches from swap on idle into free memory, ie: it doesn't force 

the fact that there is free memory is ... strange. IN principle, Linux
keeps almost no memory free (except some emergency buffers) so that
things you swap in prematurely will BY DEFINITION go at the expense of
other things that could be there.... 

also, they take up seek time (5 to 10 msec), so if you were to read
something else at the time you get additional latency.

so the conclusion that it only can be better (if you don't care about
power) is imo premature.

Sure it still can be worth it, but it for sure is NOT an automatic
thing.

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Indan Zupancic]

On Fri, July 27, 2007 22:34, Arjan van de Ven wrote:
> On Fri, July 27, 2007 21:43, Al Boldi wrote:
>> IMHO, what everybody agrees on, is that swap-prefetch has a positive effect
>> in some cases, and nobody can prove an adverse effect (excluding power
>> consumption).  The reason for this positive effect is also crystal clear:
>> It prefetches from swap on idle into free memory, ie: it doesn't force
>
> the fact that there is free memory is ... strange. IN principle, Linux
> keeps almost no memory free (except some emergency buffers) so that
> things you swap in prematurely will BY DEFINITION go at the expense of
> other things that could be there....

It's not strange, the use case here is if something memory hungry process
is shut down it leaves behind a lot of free memory. Having things swapped
out while there's free memory is strange, so swap prefetch fills it up again.

> also, they take up seek time (5 to 10 msec), so if you were to read
> something else at the time you get additional latency.

If there's other disk activity swap prefetch shouldn't do much, so this isn't
really true.

>> Conclusion:  Either prove swap-prefetch is broken, or get this merged quick.

There are a whole lot of other requirements too than that it isn't broken (of
which most are fulfilled, but anyway). One reason could be that there's a
better solution out there for the problem swap prefetch tries to solve. That
said, as swap prefetch is here now for a while and that other solution not it's
not such a great argument.

Personally I think that a more generic solution would be better, one that
prefetches the lastly evicted pages back in, not favouring either of swap or
file data, like swap prefetch does now.

Greetings,

Indan

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Arjan van de Ven]

On Fri, 2007-07-27 at 23:51 +0200, Indan Zupanci
> > also, they take up seek time (5 to 10 msec), so if you were to read
> > something else at the time you get additional latency.
> 
> If there's other disk activity swap prefetch shouldn't do much, so this isn't
> really true.

how do you know there will be other activity? You start the IO and that
basically blacks out the disk for 5 to 10 ms. If the "real" IO gets
submitted in that time you add latency. You cannot predict that IO
happening or not happening.


-- 
if you want to mail me at work (you don't), use arjan (at) linux.intel.com
Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Indan Zupancic]

On Sat, July 28, 2007 00:06, Arjan van de Ven wrote:
> On Fri, 2007-07-27 at 23:51 +0200, Indan Zupanci
>> > also, they take up seek time (5 to 10 msec), so if you were to read
>> > something else at the time you get additional latency.
>>
>> If there's other disk activity swap prefetch shouldn't do much, so this isn't
>> really true.
>
> how do you know there will be other activity? You start the IO and that
> basically blacks out the disk for 5 to 10 ms. If the "real" IO gets
> submitted in that time you add latency. You cannot predict that IO
> happening or not happening.

Ah, in that way. Yes, you right about that (though NCQ might help then?),
but that's true for all disk activity. Though I think swap prefetch didn't want
to run when there was CPU activity, so that would reduce the chance that
new IO is submitted right at that moment. I think in practice this isn't worth
worrying about, the real issue is the extra disk activity in the first place.

Greetings,

Indan

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[grundig]

El Fri, 27 Jul 2007 15:06:14 -0700, Arjan van de Ven <arjan@infradead.org> escribió:

> how do you know there will be other activity? You start the IO and that
> basically blacks out the disk for 5 to 10 ms. If the "real" IO gets
> submitted in that time you add latency. You cannot predict that IO
> happening or not happening.

If there hasn't be much IO for some time, it looks quite reasonable to expect
that there won't be more in the near future. As most of heuristics can fail, but
then this is a feature mostly for desktops, not servers.

There's an old saying that says something like "an open source project starts
dying when new people can't participate in the project no matter how hard
they try". It's hard to understand why there's so many people opposing to
this when other more controversial features are merged much faster, (like, fe. 
the UIO driver framework). 

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Arjan van de Ven]

On Sat, 2007-07-28 at 01:34 +0200, grundig wrote:
> El Fri, 27 Jul 2007 15:06:14 -0700, Arjan van de Ven <arjan@infradead.org> escribió:
> 
> > how do you know there will be other activity? You start the IO and that
> > basically blacks out the disk for 5 to 10 ms. If the "real" IO gets
> > submitted in that time you add latency. You cannot predict that IO
> > happening or not happening.
> 
> If there hasn't be much IO for some time, it looks quite reasonable to expect
> that there won't be more in the near future. As most of heuristics can fail

exactly this was my point: just saying "there are no downsides" isn't
true.

> There's an old saying that says something like "an open source project starts
> dying when new people can't participate in the project no matter how hard
> they try". It's hard to understand why there's so many people opposing to
> this when other more controversial features are merged much faster, (like, fe. 
> the UIO driver framework). 

I'm not opposing this or cheering for it. I'm opposing blindly saying
"there are no downsides". This needs showing with data at minimum, and
my reading of this saga seems to suggest data is the bit that is lacking
from the start...

-- 
if you want to mail me at work (you don't), use arjan (at) linux.intel.com
Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Indan Zupancic]

On Sat, July 28, 2007 01:34, grundig wrote:
> El Fri, 27 Jul 2007 15:06:14 -0700, Arjan van de Ven <arjan@infradead.org> escribi�:
>
>> how do you know there will be other activity? You start the IO and that
>> basically blacks out the disk for 5 to 10 ms. If the "real" IO gets
>> submitted in that time you add latency. You cannot predict that IO
>> happening or not happening.
>
> If there hasn't be much IO for some time, it looks quite reasonable to expect
> that there won't be more in the near future.

Good argument.

> As most of heuristics can fail, but
> then this is a feature mostly for desktops, not servers.

Bad argument. It doesn't matter for who the feature is intended, it matter
what it does and if it does it well or not. In this case, prefetching swap without
disturbing anything else.

> There's an old saying that says something like "an open source project starts
> dying when new people can't participate in the project no matter how hard
> they try". It's hard to understand why there's so many people opposing to
> this when other more controversial features are merged much faster, (like, fe.
> the UIO driver framework).

Could people please stop this emotional crap non-argumentation? At best it reduces
the chance of swap-prefetch to be merged.

Perhaps one of the reasons is that this is core kernel code. And that it isn't a new
feature, but a performance improvement with doubtful trade-offs. The problem
statement isn't clear either. It seems like a natural enhancement, but is that enough
reason to merge it? Maybe, maybe not. But if slow swap-in is the problem, shouldn't
that be fixed instead of bypassed?

Yes, there are people that say that it works for them, but of those a lot claim
updatedb damage is fixed by it too, while that can't be true. And how many of those
people did test swap prefetch stand-alone? The ck kernel has other mm patches too,
perhaps those are the real goodies...

And there don't seem to be many people opposing swap prefetch either. A bunch seem
in favour of it, and others seem unconvinced.

Me, I don't know if it should be merged or not, it solves one very specific workload, and
nothing else (swap is used, and memory becomes free which won't be used in the near
future). All in all it seems good, but doubtful, and when in doubt, don't merge.

Greetings,

Indan

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[grundig]

El Sat, 28 Jul 2007 02:03:19 +0200 (CEST), "Indan Zupancic" <indan@nul.nu> escribió:

> Perhaps one of the reasons is that this is core kernel code. And that it isn't a new
> feature, but a performance improvement with doubtful trade-offs. The problem
> statement isn't clear either. It seems like a natural enhancement, but is that enough
> reason to merge it? Maybe, maybe not. But if slow swap-in is the problem, shouldn't
> that be fixed instead of bypassed?


The problem AFAIK it's clear: Swap in is slow by nature because it's random. Well, it "looks"
like it should be random. Now that would be a interesting thing to analyze: locality of the
swapped-in stuff for different worloads - if the swapped-in stuff has some locality, then
it certainly should possible to fix the swapin code to do some kind of "swapahead". But if
the stuff is swapped in completely randomly, only stuff like swap prefetch is going to help.

The good thing about swap prefetch is that it fixes the problem today without forbidding
future improvements to the VM.

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Chris Snook]

Al Boldi wrote:
> People wrote:
>>>> I believe the users who say their apps really do get paged back in
>>>> though, so suspect that's not the case.
>>> Stopping the bush-circumference beating, I do not. -ck (and gentoo) have
>>> this massive Calimero thing going among their users where people are
>>> much less interested in technology than in how the nasty big kernel
>>> meanies are keeping them down (*).
>> I think the problem is elsewhere. Users don't say: "My apps get paged
>> back in." They say: "My system is more responsive". They really don't
>> care *why* the reaction to a mouse click that takes three seconds with
>> a mainline kernel is instantaneous with -ck. Nasty big kernel meanies,
>> OTOH, want to understand *why* a patch helps in order to decide whether
>> it is really a good idea to merge it. So you've got a bunch of patches
>> (aka -ck) which visibly improve the overall responsiveness of a desktop
>> system, but apparently no one can conclusively explain why or how they
>> achieve that, and therefore they cannot be merged into mainline.
>>
>> I don't have a solution to that dilemma either.
> 
> IMHO, what everybody agrees on, is that swap-prefetch has a positive effect 
> in some cases, and nobody can prove an adverse effect (excluding power 
> consumption).  The reason for this positive effect is also crystal clear:  
> It prefetches from swap on idle into free memory, ie: it doesn't force 
> anybody out, and they are the first to be dropped without further swap-out, 
> which sounds really smart.
> 
> Conclusion:  Either prove swap-prefetch is broken, or get this merged quick.

If you can't prove why it helps and doesn't hurt, then it's a hack, by 
definition.  Behind any performance hack is some fundamental truth that can be 
exploited to greater effect if we reason about it.  So let's reason about it. 
I'll start.

Resource size has been outpacing processing latency since the dawn of time. 
Disks get bigger much faster than seek times shrink.  Main memory and cache keep 
growing, while single-threaded processing speed has nearly ground to a halt.

In the old days, it made lots of sense to manage resource allocation in pages 
and blocks.  In the past few years, we started reserving blocks in ext3 
automatically because it saves more in seek time than it costs in disk space. 
Now we're taking preallocation and antifragmentation to the next level with 
extent-based allocation in ext4.

Well, we're still using bitmap-style allocation for pages, and the prefetch-less 
swap mechanism adheres to this design as well.  Maybe it's time to start 
thinking about memory in a somewhat more extent-like fashion.

With swap prefetch, we're only optimizing the case when the box isn't loaded and 
there's RAM free, but we're not optimizing the case when the box is heavily 
loaded and we need for RAM to be free.  This is a complete reversal of sane 
development priorities.  If swap batching is an optimization at all (and we have 
empirical evidence that it is) then it should also be an optimization to swap 
out chunks of pages when we need to free memory.

So, how do we go about this grouping?  I suggest that if we keep per-VMA 
reference/fault/dirty statistics, we can tell which logically distinct chunks of 
memory are being regularly used.  This would also us to apply different page 
replacement policies to chunks of memory that are being used in different fashions.

With such statistics, we could then page out VMAs in 2MB chunks when we're under 
memory pressure, also giving us the option of transparently paging them back in 
to hugepages when we have the memory free, once anonymous hugepage support is in 
place.

I'm inclined to view swap prefetch as a successful scientific experiment, and 
use that data to inform a more reasoned engineering effort.  If we can design 
something intelligent which happens to behave more or less like swap prefetch 
does under the circumstances where swap prefetch helps, and does something else 
smart under the circumstances where swap prefetch makes no discernable 
difference, it'll be a much bigger improvement.

Because we cannot prove why the existing patch helps, we cannot say what impact 
it will have when things like virtualization and solid state drives radically 
change the coefficients of the equation we have not solved.  Providing a sysctl 
to turn off a misbehaving feature is a poor substitute for doing it right the 
first time, and leaving it off by default will ensure that it only gets used by 
the handful of people who know enough to rebuild with the patch anyway.

Let's talk about how we can make page replacement smarter, so it naturally 
accomplishes what swap prefetch accomplishes, as part of a design we can reason 
about.

CC-ing linux-mm, since that's where I think we should take this next.

	-- Chris

How can we make page replacement smarter (was: swap-prefetch)

[Al Boldi]

Chris Snook wrote:
> Resource size has been outpacing processing latency since the dawn of
> time. Disks get bigger much faster than seek times shrink.  Main memory
> and cache keep growing, while single-threaded processing speed has nearly
> ground to a halt.
>
> In the old days, it made lots of sense to manage resource allocation in
> pages and blocks.  In the past few years, we started reserving blocks in
> ext3 automatically because it saves more in seek time than it costs in
> disk space. Now we're taking preallocation and antifragmentation to the
> next level with extent-based allocation in ext4.
>
> Well, we're still using bitmap-style allocation for pages, and the
> prefetch-less swap mechanism adheres to this design as well.  Maybe it's
> time to start thinking about memory in a somewhat more extent-like
> fashion.
>
> With swap prefetch, we're only optimizing the case when the box isn't
> loaded and there's RAM free, but we're not optimizing the case when the
> box is heavily loaded and we need for RAM to be free.  This is a complete
> reversal of sane development priorities.  If swap batching is an
> optimization at all (and we have empirical evidence that it is) then it
> should also be an optimization to swap out chunks of pages when we need to
> free memory.
>
> So, how do we go about this grouping?  I suggest that if we keep per-VMA
> reference/fault/dirty statistics, we can tell which logically distinct
> chunks of memory are being regularly used.  This would also us to apply
> different page replacement policies to chunks of memory that are being
> used in different fashions.
>
> With such statistics, we could then page out VMAs in 2MB chunks when we're
> under memory pressure, also giving us the option of transparently paging
> them back in to hugepages when we have the memory free, once anonymous
> hugepage support is in place.
>
> I'm inclined to view swap prefetch as a successful scientific experiment,
> and use that data to inform a more reasoned engineering effort.  If we can
> design something intelligent which happens to behave more or less like
> swap prefetch does under the circumstances where swap prefetch helps, and
> does something else smart under the circumstances where swap prefetch
> makes no discernable difference, it'll be a much bigger improvement.
>
> Because we cannot prove why the existing patch helps, we cannot say what
> impact it will have when things like virtualization and solid state drives
> radically change the coefficients of the equation we have not solved. 
> Providing a sysctl to turn off a misbehaving feature is a poor substitute
> for doing it right the first time, and leaving it off by default will
> ensure that it only gets used by the handful of people who know enough to
> rebuild with the patch anyway.
>
> Let's talk about how we can make page replacement smarter, so it naturally
> accomplishes what swap prefetch accomplishes, as part of a design we can
> reason about.
>
> CC-ing linux-mm, since that's where I think we should take this next.

Good idea, but unless we understand the problems involved, we are bound to 
repeat it.  So my first question would be:  Why is swap-in so slow?

As I have posted in other threads, swap-in of consecutive pages suffers a 2x 
slowdown wrt swap-out, whereas swap-in of random pages suffers over 6x 
slowdown.

Because it is hard to quantify the expected swap-in speed for random pages, 
let's first tackle the swap-in of consecutive pages, which should be at 
least as fast as swap-out.  So again, why is swap-in so slow?

Once we understand this problem, we may be able to suggest a smart 
improvement.


Thanks!

--
Al

Re: How can we make page replacement smarter (was: swap-prefetch)

[Chris Snook]

Al Boldi wrote:
> Chris Snook wrote:
>> Resource size has been outpacing processing latency since the dawn of
>> time. Disks get bigger much faster than seek times shrink.  Main memory
>> and cache keep growing, while single-threaded processing speed has nearly
>> ground to a halt.
>>
>> In the old days, it made lots of sense to manage resource allocation in
>> pages and blocks.  In the past few years, we started reserving blocks in
>> ext3 automatically because it saves more in seek time than it costs in
>> disk space. Now we're taking preallocation and antifragmentation to the
>> next level with extent-based allocation in ext4.
>>
>> Well, we're still using bitmap-style allocation for pages, and the
>> prefetch-less swap mechanism adheres to this design as well.  Maybe it's
>> time to start thinking about memory in a somewhat more extent-like
>> fashion.
>>
>> With swap prefetch, we're only optimizing the case when the box isn't
>> loaded and there's RAM free, but we're not optimizing the case when the
>> box is heavily loaded and we need for RAM to be free.  This is a complete
>> reversal of sane development priorities.  If swap batching is an
>> optimization at all (and we have empirical evidence that it is) then it
>> should also be an optimization to swap out chunks of pages when we need to
>> free memory.
>>
>> So, how do we go about this grouping?  I suggest that if we keep per-VMA
>> reference/fault/dirty statistics, we can tell which logically distinct
>> chunks of memory are being regularly used.  This would also us to apply
>> different page replacement policies to chunks of memory that are being
>> used in different fashions.
>>
>> With such statistics, we could then page out VMAs in 2MB chunks when we're
>> under memory pressure, also giving us the option of transparently paging
>> them back in to hugepages when we have the memory free, once anonymous
>> hugepage support is in place.
>>
>> I'm inclined to view swap prefetch as a successful scientific experiment,
>> and use that data to inform a more reasoned engineering effort.  If we can
>> design something intelligent which happens to behave more or less like
>> swap prefetch does under the circumstances where swap prefetch helps, and
>> does something else smart under the circumstances where swap prefetch
>> makes no discernable difference, it'll be a much bigger improvement.
>>
>> Because we cannot prove why the existing patch helps, we cannot say what
>> impact it will have when things like virtualization and solid state drives
>> radically change the coefficients of the equation we have not solved. 
>> Providing a sysctl to turn off a misbehaving feature is a poor substitute
>> for doing it right the first time, and leaving it off by default will
>> ensure that it only gets used by the handful of people who know enough to
>> rebuild with the patch anyway.
>>
>> Let's talk about how we can make page replacement smarter, so it naturally
>> accomplishes what swap prefetch accomplishes, as part of a design we can
>> reason about.
>>
>> CC-ing linux-mm, since that's where I think we should take this next.
> 
> Good idea, but unless we understand the problems involved, we are bound to 
> repeat it.  So my first question would be:  Why is swap-in so slow?
> 
> As I have posted in other threads, swap-in of consecutive pages suffers a 2x 
> slowdown wrt swap-out, whereas swap-in of random pages suffers over 6x 
> slowdown.
> 
> Because it is hard to quantify the expected swap-in speed for random pages, 
> let's first tackle the swap-in of consecutive pages, which should be at 
> least as fast as swap-out.  So again, why is swap-in so slow?

If I'm writing 20 pages to swap, I can find a suitable chunk of swap and 
write them all in one place.  If I'm reading 20 pages from swap, they 
could be anywhere.  Also, writes get buffered at one or more layers of 
hardware.  At best, reads can be read-ahead and cached, which is why 
sequential swap-in sucks less.  On-demand reads are as expensive as I/O 
can get.

> Once we understand this problem, we may be able to suggest a smart 
> improvement.

There are lots of page replacement schemes that optimize for different 
access patterns, and they all suck at certain other access patterns.  We 
tweak our behavior slightly based on fadvise and madvise hints, but most 
of the memory we're managing is an opaque mass.  With more statistics, 
we could do a better job of managing chunks of unhinted memory with 
disparate access patterns.  Of course, this imposes overhead.  I 
suggested VMA granularity because a VMA represents a logically distinct 
piece of address space, though this may not be suitable for shared mappings.

	-- Chris

Re: How can we make page replacement smarter (was: swap-prefetch)

[Al Boldi]

Chris Snook wrote:
> Al Boldi wrote:
> > Because it is hard to quantify the expected swap-in speed for random
> > pages, let's first tackle the swap-in of consecutive pages, which should
> > be at least as fast as swap-out.  So again, why is swap-in so slow?
>
> If I'm writing 20 pages to swap, I can find a suitable chunk of swap and
> write them all in one place.  If I'm reading 20 pages from swap, they
> could be anywhere.  Also, writes get buffered at one or more layers of
> hardware.

Ok, this explains swap-in of random pages.  Makes sense, but it doesn't 
explain the awful tmpfs performance degradation of consecutive read-in runs 
from swap, which should have at least stayed constant

> At best, reads can be read-ahead and cached, which is why
> sequential swap-in sucks less.  On-demand reads are as expensive as I/O
> can get.

Which means that it should be at least as fast as swap-out, even faster 
because write to disk is usually slower than read on modern disks.  But 
linux currently shows a distinct 2x slowdown for sequential swap-in wrt 
swap-out.  And to prove this point, just try suspend to disk where you can 
see sequential swap-out being reported at about twice the speed of 
sequential swap-in on resume.  Why is that?


Thanks!

--
Al

Re: How can we make page replacement smarter (was: swap-prefetch)

[Rik van Riel]

Al Boldi wrote:
> Chris Snook wrote:

>> At best, reads can be read-ahead and cached, which is why
>> sequential swap-in sucks less.  On-demand reads are as expensive as I/O
>> can get.
> 
> Which means that it should be at least as fast as swap-out, even faster 
> because write to disk is usually slower than read on modern disks.  But 
> linux currently shows a distinct 2x slowdown for sequential swap-in wrt 
> swap-out. 

That's because writes are faster than reads in moderate
quantities.

The disk caches writes, allowing the OS to write a whole
bunch of data into the disk cache and the disk can optimize
the IO a bit internally.

The same optimization is not possible for reads.

-- 
Politics is the struggle between those who want to make their country
the best in the world, and those who believe it already is.  Each group
calls the other unpatriotic.

Re: How can we make page replacement smarter

[Rik van Riel]

[-- Attachment #1: Type: text/plain, Size: 1817 bytes --]

Al Boldi wrote:

> Good idea, but unless we understand the problems involved, we are bound to 
> repeat it.  So my first question would be:  Why is swap-in so slow?
> 
> As I have posted in other threads, swap-in of consecutive pages suffers a 2x 
> slowdown wrt swap-out, whereas swap-in of random pages suffers over 6x 
> slowdown.
> 
> Because it is hard to quantify the expected swap-in speed for random pages, 
> let's first tackle the swap-in of consecutive pages, which should be at 
> least as fast as swap-out.  So again, why is swap-in so slow?

I suspect that this is a locality of reference issue.

Anonymous memory can get jumbled up by repeated free and
malloc cycles of many smaller objects.  The amount of
anonymous memory is often smaller than or roughly the same
size as system memory.

Locality of refenence to anonymous memory tends to be
temporal in nature, with the same sets of pages being
accessed over and over again.

Files are different.  File content tends to be grouped
in large related chunks, both logically in the file and
on disk.  Generally there is a lot more file data on a
system than what fits in memory.

Locality of reference to file data tends to be spatial
in nature, with one file access leading up to the system
accessing "nearby" data.  The data is not necessarily
touched again any time soon.

> Once we understand this problem, we may be able to suggest a smart 
> improvement.

Like the one on http://linux-mm.org/PageoutFailureModes ?

I have the LRU lists split and am working on getting SEQ
replacement implemented for the anonymous pages.

The most recent (untested) patches are attached.

-- 
Politics is the struggle between those who want to make their country
the best in the world, and those who believe it already is.  Each group
calls the other unpatriotic.

[-- Attachment #2: linux-2.6.22-vmsplit.patch --]
[-- Type: text/x-patch, Size: 50691 bytes --]

--- linux-2.6.21.noarch/drivers/base/node.c.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/drivers/base/node.c	2007-07-23 11:42:52.000000000 -0400
@@ -44,33 +44,37 @@ static ssize_t node_read_meminfo(struct 
 	si_meminfo_node(&i, nid);
 
 	n = sprintf(buf, "\n"
-		       "Node %d MemTotal:     %8lu kB\n"
-		       "Node %d MemFree:      %8lu kB\n"
-		       "Node %d MemUsed:      %8lu kB\n"
-		       "Node %d Active:       %8lu kB\n"
-		       "Node %d Inactive:     %8lu kB\n"
+		       "Node %d MemTotal:       %8lu kB\n"
+		       "Node %d MemFree:        %8lu kB\n"
+		       "Node %d MemUsed:        %8lu kB\n"
+		       "Node %d Active(anon):   %8lu kB\n"
+		       "Node %d Inactive(anon): %8lu kB\n"
+		       "Node %d Active(file):   %8lu kB\n"
+		       "Node %d Inactive(file): %8lu kB\n"
 #ifdef CONFIG_HIGHMEM
-		       "Node %d HighTotal:    %8lu kB\n"
-		       "Node %d HighFree:     %8lu kB\n"
-		       "Node %d LowTotal:     %8lu kB\n"
-		       "Node %d LowFree:      %8lu kB\n"
+		       "Node %d HighTotal:      %8lu kB\n"
+		       "Node %d HighFree:       %8lu kB\n"
+		       "Node %d LowTotal:       %8lu kB\n"
+		       "Node %d LowFree:        %8lu kB\n"
 #endif
-		       "Node %d Dirty:        %8lu kB\n"
-		       "Node %d Writeback:    %8lu kB\n"
-		       "Node %d FilePages:    %8lu kB\n"
-		       "Node %d Mapped:       %8lu kB\n"
-		       "Node %d AnonPages:    %8lu kB\n"
-		       "Node %d PageTables:   %8lu kB\n"
-		       "Node %d NFS_Unstable: %8lu kB\n"
-		       "Node %d Bounce:       %8lu kB\n"
-		       "Node %d Slab:         %8lu kB\n"
-		       "Node %d SReclaimable: %8lu kB\n"
-		       "Node %d SUnreclaim:   %8lu kB\n",
+		       "Node %d Dirty:          %8lu kB\n"
+		       "Node %d Writeback:      %8lu kB\n"
+		       "Node %d FilePages:      %8lu kB\n"
+		       "Node %d Mapped:         %8lu kB\n"
+		       "Node %d AnonPages:      %8lu kB\n"
+		       "Node %d PageTables:     %8lu kB\n"
+		       "Node %d NFS_Unstable:   %8lu kB\n"
+		       "Node %d Bounce:         %8lu kB\n"
+		       "Node %d Slab:           %8lu kB\n"
+		       "Node %d SReclaimable:   %8lu kB\n"
+		       "Node %d SUnreclaim:     %8lu kB\n",
 		       nid, K(i.totalram),
 		       nid, K(i.freeram),
 		       nid, K(i.totalram - i.freeram),
-		       nid, node_page_state(nid, NR_ACTIVE),
-		       nid, node_page_state(nid, NR_INACTIVE),
+		       nid, node_page_state(nid, NR_ACTIVE_ANON),
+		       nid, node_page_state(nid, NR_INACTIVE_ANON),
+		       nid, node_page_state(nid, NR_ACTIVE_FILE),
+		       nid, node_page_state(nid, NR_INACTIVE_FILE),
 #ifdef CONFIG_HIGHMEM
 		       nid, K(i.totalhigh),
 		       nid, K(i.freehigh),
--- linux-2.6.21.noarch/fs/proc/proc_misc.c.vmsplit	2007-07-05 12:06:14.000000000 -0400
+++ linux-2.6.21.noarch/fs/proc/proc_misc.c	2007-07-23 11:42:52.000000000 -0400
@@ -146,43 +146,47 @@ static int meminfo_read_proc(char *page,
 	 * Tagged format, for easy grepping and expansion.
 	 */
 	len = sprintf(page,
-		"MemTotal:     %8lu kB\n"
-		"MemFree:      %8lu kB\n"
-		"Buffers:      %8lu kB\n"
-		"Cached:       %8lu kB\n"
-		"SwapCached:   %8lu kB\n"
-		"Active:       %8lu kB\n"
-		"Inactive:     %8lu kB\n"
+		"MemTotal:       %8lu kB\n"
+		"MemFree:        %8lu kB\n"
+		"Buffers:        %8lu kB\n"
+		"Cached:         %8lu kB\n"
+		"SwapCached:     %8lu kB\n"
+		"Active(anon):   %8lu kB\n"
+		"Inactive(anon): %8lu kB\n"
+		"Active(file):   %8lu kB\n"
+		"Inactive(file): %8lu kB\n"
 #ifdef CONFIG_HIGHMEM
-		"HighTotal:    %8lu kB\n"
-		"HighFree:     %8lu kB\n"
-		"LowTotal:     %8lu kB\n"
-		"LowFree:      %8lu kB\n"
-#endif
-		"SwapTotal:    %8lu kB\n"
-		"SwapFree:     %8lu kB\n"
-		"Dirty:        %8lu kB\n"
-		"Writeback:    %8lu kB\n"
-		"AnonPages:    %8lu kB\n"
-		"Mapped:       %8lu kB\n"
-		"Slab:         %8lu kB\n"
-		"SReclaimable: %8lu kB\n"
-		"SUnreclaim:   %8lu kB\n"
-		"PageTables:   %8lu kB\n"
-		"NFS_Unstable: %8lu kB\n"
-		"Bounce:       %8lu kB\n"
-		"CommitLimit:  %8lu kB\n"
-		"Committed_AS: %8lu kB\n"
-		"VmallocTotal: %8lu kB\n"
-		"VmallocUsed:  %8lu kB\n"
-		"VmallocChunk: %8lu kB\n",
+		"HighTotal:      %8lu kB\n"
+		"HighFree:       %8lu kB\n"
+		"LowTotal:       %8lu kB\n"
+		"LowFree:        %8lu kB\n"
+#endif
+		"SwapTotal:      %8lu kB\n"
+		"SwapFree:       %8lu kB\n"
+		"Dirty:          %8lu kB\n"
+		"Writeback:      %8lu kB\n"
+		"AnonPages:      %8lu kB\n"
+		"Mapped:         %8lu kB\n"
+		"Slab:           %8lu kB\n"
+		"SReclaimable:   %8lu kB\n"
+		"SUnreclaim:     %8lu kB\n"
+		"PageTables:     %8lu kB\n"
+		"NFS_Unstable:   %8lu kB\n"
+		"Bounce:         %8lu kB\n"
+		"CommitLimit:    %8lu kB\n"
+		"Committed_AS:   %8lu kB\n"
+		"VmallocTotal:   %8lu kB\n"
+		"VmallocUsed:    %8lu kB\n"
+		"VmallocChunk:   %8lu kB\n",
 		K(i.totalram),
 		K(i.freeram),
 		K(i.bufferram),
 		K(cached),
 		K(total_swapcache_pages),
-		K(global_page_state(NR_ACTIVE)),
-		K(global_page_state(NR_INACTIVE)),
+		K(global_page_state(NR_ACTIVE_ANON)),
+		K(global_page_state(NR_INACTIVE_ANON)),
+		K(global_page_state(NR_ACTIVE_FILE)),
+		K(global_page_state(NR_INACTIVE_FILE)),
 #ifdef CONFIG_HIGHMEM
 		K(i.totalhigh),
 		K(i.freehigh),
--- linux-2.6.21.noarch/fs/exec.c.vmsplit	2007-07-05 12:06:22.000000000 -0400
+++ linux-2.6.21.noarch/fs/exec.c	2007-07-23 11:42:52.000000000 -0400
@@ -327,7 +327,7 @@ void install_arg_page(struct vm_area_str
 		goto out;
 	}
 	inc_mm_counter(mm, anon_rss);
-	lru_cache_add_active(page);
+	lru_cache_add_active_anon(page);
 	set_pte_at(mm, address, pte, pte_mkdirty(pte_mkwrite(mk_pte(
 					page, vma->vm_page_prot))));
 	page_add_new_anon_rmap(page, vma, address);
--- linux-2.6.21.noarch/fs/mpage.c.vmsplit	2007-07-05 12:06:14.000000000 -0400
+++ linux-2.6.21.noarch/fs/mpage.c	2007-07-23 11:42:52.000000000 -0400
@@ -406,12 +406,12 @@ mpage_readpages(struct address_space *ma
 					&first_logical_block,
 					get_block);
 			if (!pagevec_add(&lru_pvec, page))
-				__pagevec_lru_add(&lru_pvec);
+				__pagevec_lru_add_file(&lru_pvec);
 		} else {
 			page_cache_release(page);
 		}
 	}
-	pagevec_lru_add(&lru_pvec);
+	pagevec_lru_add_file(&lru_pvec);
 	BUG_ON(!list_empty(pages));
 	if (bio)
 		mpage_bio_submit(READ, bio);
--- linux-2.6.21.noarch/include/linux/mm.h.vmsplit	2007-07-05 12:06:24.000000000 -0400
+++ linux-2.6.21.noarch/include/linux/mm.h	2007-07-23 11:42:52.000000000 -0400
@@ -617,6 +617,24 @@ static inline int PageAnon(struct page *
 	return ((unsigned long)page->mapping & PAGE_MAPPING_ANON) != 0;
 }
 
+/* Returns true if this page is anonymous, tmpfs or otherwise swap backed. */
+extern const struct address_space_operations shmem_aops;
+static inline int page_anon(struct page *page)
+{
+	struct address_space *mapping = page_mapping(page);
+
+	if (PageAnon(page) || PageSwapCache(page))
+		return 1;
+	if (mapping->a_ops == &shmem_aops)
+		return 1;
+	/* Should ramfs pages go onto an mlocked list instead? */
+	if ((unlikely(mapping->a_ops->writepage == NULL)))
+		return 1;
+
+	/* The page is page cache backed by a normal filesystem. */
+	return 0;
+}
+
 /*
  * Return the pagecache index of the passed page.  Regular pagecache pages
  * use ->index whereas swapcache pages use ->private
--- linux-2.6.21.noarch/include/linux/mm_inline.h.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/include/linux/mm_inline.h	2007-07-23 11:42:52.000000000 -0400
@@ -1,29 +1,57 @@
 static inline void
-add_page_to_active_list(struct zone *zone, struct page *page)
+add_page_to_active_anon_list(struct zone *zone, struct page *page)
 {
-	list_add(&page->lru, &zone->active_list);
-	__inc_zone_state(zone, NR_ACTIVE);
+	list_add(&page->lru, &zone->active_anon_list);
+	__inc_zone_state(zone, NR_ACTIVE_ANON);
 }
 
 static inline void
-add_page_to_inactive_list(struct zone *zone, struct page *page)
+add_page_to_inactive_anon_list(struct zone *zone, struct page *page)
 {
-	list_add(&page->lru, &zone->inactive_list);
-	__inc_zone_state(zone, NR_INACTIVE);
+	list_add(&page->lru, &zone->inactive_anon_list);
+	__inc_zone_state(zone, NR_INACTIVE_ANON);
 }
 
 static inline void
-del_page_from_active_list(struct zone *zone, struct page *page)
+del_page_from_active_anon_list(struct zone *zone, struct page *page)
 {
 	list_del(&page->lru);
-	__dec_zone_state(zone, NR_ACTIVE);
+	__dec_zone_state(zone, NR_ACTIVE_ANON);
 }
 
 static inline void
-del_page_from_inactive_list(struct zone *zone, struct page *page)
+del_page_from_inactive_anon_list(struct zone *zone, struct page *page)
 {
 	list_del(&page->lru);
-	__dec_zone_state(zone, NR_INACTIVE);
+	__dec_zone_state(zone, NR_INACTIVE_ANON);
+}
+
+static inline void
+add_page_to_active_file_list(struct zone *zone, struct page *page)
+{
+	list_add(&page->lru, &zone->active_file_list);
+	__inc_zone_state(zone, NR_ACTIVE_FILE);
+}
+
+static inline void
+add_page_to_inactive_file_list(struct zone *zone, struct page *page)
+{
+	list_add(&page->lru, &zone->inactive_file_list);
+	__inc_zone_state(zone, NR_INACTIVE_FILE);
+}
+
+static inline void
+del_page_from_active_file_list(struct zone *zone, struct page *page)
+{
+	list_del(&page->lru);
+	__dec_zone_state(zone, NR_ACTIVE_FILE);
+}
+
+static inline void
+del_page_from_inactive_file_list(struct zone *zone, struct page *page)
+{
+	list_del(&page->lru);
+	__dec_zone_state(zone, NR_INACTIVE_FILE);
 }
 
 static inline void
@@ -32,9 +60,15 @@ del_page_from_lru(struct zone *zone, str
 	list_del(&page->lru);
 	if (PageActive(page)) {
 		__ClearPageActive(page);
-		__dec_zone_state(zone, NR_ACTIVE);
+		if (page_anon(page))
+			__dec_zone_state(zone, NR_ACTIVE_ANON);
+		else
+			__dec_zone_state(zone, NR_ACTIVE_FILE);
 	} else {
-		__dec_zone_state(zone, NR_INACTIVE);
+		if (page_anon(page))
+			__dec_zone_state(zone, NR_INACTIVE_ANON);
+		else
+			__dec_zone_state(zone, NR_INACTIVE_FILE);
 	}
 }
 
--- linux-2.6.21.noarch/include/linux/mmzone.h.vmsplit	2007-07-05 12:06:18.000000000 -0400
+++ linux-2.6.21.noarch/include/linux/mmzone.h	2007-07-23 11:43:35.000000000 -0400
@@ -49,15 +49,17 @@ struct zone_padding {
 enum zone_stat_item {
 	/* First 128 byte cacheline (assuming 64 bit words) */
 	NR_FREE_PAGES,
-	NR_INACTIVE,
-	NR_ACTIVE,
+	NR_ACTIVE_ANON,
+	NR_INACTIVE_ANON,
+	NR_ACTIVE_FILE,
+	NR_INACTIVE_FILE,
 	NR_ANON_PAGES,	/* Mapped anonymous pages */
 	NR_FILE_MAPPED,	/* pagecache pages mapped into pagetables.
 			   only modified from process context */
 	NR_FILE_PAGES,
+	/* Second 128 byte cacheline */
 	NR_FILE_DIRTY,
 	NR_WRITEBACK,
-	/* Second 128 byte cacheline */
 	NR_SLAB_RECLAIMABLE,
 	NR_SLAB_UNRECLAIMABLE,
 	NR_PAGETABLE,		/* used for pagetables */
@@ -218,10 +220,18 @@ struct zone {
 
 	/* Fields commonly accessed by the page reclaim scanner */
 	spinlock_t		lru_lock;	
-	struct list_head	active_list;
-	struct list_head	inactive_list;
-	unsigned long		nr_scan_active;
-	unsigned long		nr_scan_inactive;
+	struct list_head	active_anon_list;
+	struct list_head	inactive_anon_list;
+	struct list_head	active_file_list;
+	struct list_head	inactive_file_list;
+	unsigned long		nr_scan_active_anon;
+	unsigned long		nr_scan_inactive_anon;
+	unsigned long		nr_scan_active_file;
+	unsigned long		nr_scan_inactive_file;
+	unsigned long		recent_rotated_anon;
+	unsigned long		recent_rotated_file;
+	unsigned long		recent_scanned_anon;
+	unsigned long		recent_scanned_file;
 	unsigned long		pages_scanned;	   /* since last reclaim */
 	int			all_unreclaimable; /* All pages pinned */
 
--- linux-2.6.21.noarch/include/linux/pagevec.h.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/include/linux/pagevec.h	2007-07-23 11:42:52.000000000 -0400
@@ -23,8 +23,10 @@ struct pagevec {
 void __pagevec_release(struct pagevec *pvec);
 void __pagevec_release_nonlru(struct pagevec *pvec);
 void __pagevec_free(struct pagevec *pvec);
-void __pagevec_lru_add(struct pagevec *pvec);
-void __pagevec_lru_add_active(struct pagevec *pvec);
+void __pagevec_lru_add_file(struct pagevec *pvec);
+void __pagevec_lru_add_active_file(struct pagevec *pvec);
+void __pagevec_lru_add_anon(struct pagevec *pvec);
+void __pagevec_lru_add_active_anon(struct pagevec *pvec);
 void pagevec_strip(struct pagevec *pvec);
 unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
 		pgoff_t start, unsigned nr_pages);
@@ -81,10 +83,16 @@ static inline void pagevec_free(struct p
 		__pagevec_free(pvec);
 }
 
-static inline void pagevec_lru_add(struct pagevec *pvec)
+static inline void pagevec_lru_add_file(struct pagevec *pvec)
 {
 	if (pagevec_count(pvec))
-		__pagevec_lru_add(pvec);
+		__pagevec_lru_add_file(pvec);
+}
+
+static inline void pagevec_lru_add_anon(struct pagevec *pvec)
+{
+	if (pagevec_count(pvec))
+		__pagevec_lru_add_anon(pvec);
 }
 
 #endif /* _LINUX_PAGEVEC_H */
--- linux-2.6.21.noarch/include/linux/swap.h.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/include/linux/swap.h	2007-07-23 12:21:11.000000000 -0400
@@ -178,9 +178,12 @@ extern unsigned int nr_free_pagecache_pa
 
 
 /* linux/mm/swap.c */
-extern void FASTCALL(lru_cache_add(struct page *));
-extern void FASTCALL(lru_cache_add_active(struct page *));
-extern void FASTCALL(activate_page(struct page *));
+extern void FASTCALL(lru_cache_add_file(struct page *));
+extern void FASTCALL(lru_cache_add_anon(struct page *));
+extern void FASTCALL(lru_cache_add_active_file(struct page *));
+extern void FASTCALL(lru_cache_add_active_anon(struct page *));
+extern void FASTCALL(activate_page_anon(struct page *));
+extern void FASTCALL(activate_page_file(struct page *));
 extern void FASTCALL(mark_page_accessed(struct page *));
 extern void lru_add_drain(void);
 extern int lru_add_drain_all(void);
--- linux-2.6.21.noarch/mm/filemap.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/filemap.c	2007-07-23 11:42:52.000000000 -0400
@@ -462,7 +462,7 @@ int add_to_page_cache_lru(struct page *p
 {
 	int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 	if (ret == 0)
-		lru_cache_add(page);
+		lru_cache_add_file(page);
 	return ret;
 }
 
@@ -1870,7 +1870,7 @@ repeat:
 			page = *cached_page;
 			page_cache_get(page);
 			if (!pagevec_add(lru_pvec, page))
-				__pagevec_lru_add(lru_pvec);
+				__pagevec_lru_add_file(lru_pvec);
 			*cached_page = NULL;
 		}
 	}
@@ -2231,7 +2231,7 @@ zero_length_segment:
 	if (unlikely(file->f_flags & O_DIRECT) && written)
 		status = filemap_write_and_wait(mapping);
 
-	pagevec_lru_add(&lru_pvec);
+	pagevec_lru_add_file(&lru_pvec);
 	return written ? written : status;
 }
 EXPORT_SYMBOL(generic_file_buffered_write);
--- linux-2.6.21.noarch/mm/memory.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/memory.c	2007-07-23 11:42:52.000000000 -0400
@@ -1745,7 +1745,7 @@ gotten:
 		ptep_clear_flush(vma, address, page_table);
 		set_pte_at(mm, address, page_table, entry);
 		update_mmu_cache(vma, address, entry);
-		lru_cache_add_active(new_page);
+		lru_cache_add_active_anon(new_page);
 		page_add_new_anon_rmap(new_page, vma, address);
 
 		/* Free the old page.. */
@@ -2242,7 +2242,7 @@ static int do_anonymous_page(struct mm_s
 		if (!pte_none(*page_table))
 			goto release;
 		inc_mm_counter(mm, anon_rss);
-		lru_cache_add_active(page);
+		lru_cache_add_active_anon(page);
 		page_add_new_anon_rmap(page, vma, address);
 	} else {
 		/* Map the ZERO_PAGE - vm_page_prot is readonly */
@@ -2389,7 +2389,7 @@ retry:
 		set_pte_at(mm, address, page_table, entry);
 		if (anon) {
 			inc_mm_counter(mm, anon_rss);
-			lru_cache_add_active(new_page);
+			lru_cache_add_active_anon(new_page);
 			page_add_new_anon_rmap(new_page, vma, address);
 		} else {
 			inc_mm_counter(mm, file_rss);
--- linux-2.6.21.noarch/mm/migrate.c.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/mm/migrate.c	2007-07-23 11:42:52.000000000 -0400
@@ -53,10 +53,17 @@ int isolate_lru_page(struct page *page, 
 			ret = 0;
 			get_page(page);
 			ClearPageLRU(page);
-			if (PageActive(page))
-				del_page_from_active_list(zone, page);
-			else
-				del_page_from_inactive_list(zone, page);
+			if (PageActive(page)) {
+			    if (page_anon(page)) 
+				del_page_from_active_anon_list(zone, page);
+			    else
+				del_page_from_active_file_list(zone, page);
+			} else {
+			    if (page_anon(page)) 
+				del_page_from_inactive_anon_list(zone, page);
+			    else
+				del_page_from_inactive_file_list(zone, page);
+			}
 			list_add_tail(&page->lru, pagelist);
 		}
 		spin_unlock_irq(&zone->lru_lock);
@@ -89,9 +96,15 @@ static inline void move_to_lru(struct pa
 		 * the PG_active bit is off.
 		 */
 		ClearPageActive(page);
-		lru_cache_add_active(page);
+		if (page_anon(page))
+			lru_cache_add_active_anon(page);
+		else
+			lru_cache_add_active_file(page);
 	} else {
-		lru_cache_add(page);
+		if (page_anon(page))
+			lru_cache_add_anon(page);
+		else
+			lru_cache_add_file(page);
 	}
 	put_page(page);
 }
--- linux-2.6.21.noarch/mm/page_alloc.c.vmsplit	2007-07-05 12:06:25.000000000 -0400
+++ linux-2.6.21.noarch/mm/page_alloc.c	2007-07-23 11:42:52.000000000 -0400
@@ -1544,10 +1544,13 @@ void show_free_areas(void)
 		}
 	}
 
-	printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
+	printk("Active_anon:%lu active_file:%lu inactive_anon%lu\n"
+		" inactive_file:%lu dirty:%lu writeback:%lu unstable:%lu\n"
 		" free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
-		global_page_state(NR_ACTIVE),
-		global_page_state(NR_INACTIVE),
+		global_page_state(NR_ACTIVE_ANON),
+		global_page_state(NR_ACTIVE_FILE),
+		global_page_state(NR_INACTIVE_ANON),
+		global_page_state(NR_INACTIVE_FILE),
 		global_page_state(NR_FILE_DIRTY),
 		global_page_state(NR_WRITEBACK),
 		global_page_state(NR_UNSTABLE_NFS),
@@ -1570,8 +1573,10 @@ void show_free_areas(void)
 			" min:%lukB"
 			" low:%lukB"
 			" high:%lukB"
-			" active:%lukB"
-			" inactive:%lukB"
+			" active_anon:%lukB"
+			" inactive_anon:%lukB"
+			" active_file:%lukB"
+			" inactive_file:%lukB"
 			" present:%lukB"
 			" pages_scanned:%lu"
 			" all_unreclaimable? %s"
@@ -1581,8 +1586,10 @@ void show_free_areas(void)
 			K(zone->pages_min),
 			K(zone->pages_low),
 			K(zone->pages_high),
-			K(zone_page_state(zone, NR_ACTIVE)),
-			K(zone_page_state(zone, NR_INACTIVE)),
+			K(zone_page_state(zone, NR_ACTIVE_ANON)),
+			K(zone_page_state(zone, NR_INACTIVE_ANON)),
+			K(zone_page_state(zone, NR_ACTIVE_FILE)),
+			K(zone_page_state(zone, NR_INACTIVE_FILE)),
 			K(zone->present_pages),
 			zone->pages_scanned,
 			(zone->all_unreclaimable ? "yes" : "no")
@@ -2647,10 +2654,16 @@ static void __meminit free_area_init_cor
 		zone->prev_priority = DEF_PRIORITY;
 
 		zone_pcp_init(zone);
-		INIT_LIST_HEAD(&zone->active_list);
-		INIT_LIST_HEAD(&zone->inactive_list);
-		zone->nr_scan_active = 0;
-		zone->nr_scan_inactive = 0;
+		INIT_LIST_HEAD(&zone->active_anon_list);
+		INIT_LIST_HEAD(&zone->inactive_anon_list);
+		INIT_LIST_HEAD(&zone->active_file_list);
+		INIT_LIST_HEAD(&zone->inactive_file_list);
+		zone->nr_scan_active_anon = 0;
+		zone->nr_scan_inactive_anon = 0;
+		zone->nr_scan_active_file = 0;
+		zone->nr_scan_inactive_file = 0;
+		zone->recent_rotated_anon = 0;
+		zone->recent_rotated_file = 0;
 		zap_zone_vm_stats(zone);
 		atomic_set(&zone->reclaim_in_progress, 0);
 		if (!size)
--- linux-2.6.21.noarch/mm/readahead.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/readahead.c	2007-07-23 11:42:52.000000000 -0400
@@ -147,14 +147,14 @@ int read_cache_pages(struct address_spac
 		}
 		ret = filler(data, page);
 		if (!pagevec_add(&lru_pvec, page))
-			__pagevec_lru_add(&lru_pvec);
+			__pagevec_lru_add_file(&lru_pvec);
 		if (ret) {
 			put_pages_list(pages);
 			break;
 		}
 		task_io_account_read(PAGE_CACHE_SIZE);
 	}
-	pagevec_lru_add(&lru_pvec);
+	pagevec_lru_add_file(&lru_pvec);
 	return ret;
 }
 
@@ -182,11 +182,11 @@ static int read_pages(struct address_spa
 					page->index, GFP_KERNEL)) {
 			mapping->a_ops->readpage(filp, page);
 			if (!pagevec_add(&lru_pvec, page))
-				__pagevec_lru_add(&lru_pvec);
+				__pagevec_lru_add_file(&lru_pvec);
 		} else
 			page_cache_release(page);
 	}
-	pagevec_lru_add(&lru_pvec);
+	pagevec_lru_add_file(&lru_pvec);
 	ret = 0;
 out:
 	return ret;
@@ -578,6 +578,6 @@ void handle_ra_miss(struct address_space
  */
 unsigned long max_sane_readahead(unsigned long nr)
 {
-	return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE)
+	return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
 		+ node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
 }
--- linux-2.6.21.noarch/mm/shmem.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/shmem.c	2007-07-23 11:42:52.000000000 -0400
@@ -176,7 +176,7 @@ static inline void shmem_unacct_blocks(u
 }
 
 static const struct super_operations shmem_ops;
-static const struct address_space_operations shmem_aops;
+const struct address_space_operations shmem_aops;
 static const struct file_operations shmem_file_operations;
 static const struct inode_operations shmem_inode_operations;
 static const struct inode_operations shmem_dir_inode_operations;
@@ -2382,7 +2382,7 @@ static void destroy_inodecache(void)
 	kmem_cache_destroy(shmem_inode_cachep);
 }
 
-static const struct address_space_operations shmem_aops = {
+const struct address_space_operations shmem_aops = {
 	.writepage	= shmem_writepage,
 	.set_page_dirty	= __set_page_dirty_no_writeback,
 #ifdef CONFIG_TMPFS
--- linux-2.6.21.noarch/mm/swap_state.c.vmsplit	2007-04-25 23:08:32.000000000 -0400
+++ linux-2.6.21.noarch/mm/swap_state.c	2007-07-23 11:42:52.000000000 -0400
@@ -354,7 +354,7 @@ struct page *read_swap_cache_async(swp_e
 			/*
 			 * Initiate read into locked page and return.
 			 */
-			lru_cache_add_active(new_page);
+			lru_cache_add_anon(new_page);
 			swap_readpage(NULL, new_page);
 			return new_page;
 		}
--- linux-2.6.21.noarch/mm/vmscan.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/vmscan.c	2007-07-23 11:56:50.000000000 -0400
@@ -66,6 +66,9 @@ struct scan_control {
 	int swappiness;
 
 	int all_unreclaimable;
+
+	/* The number of pages moved to the active list this pass. */
+	int activated;
 };
 
 /*
@@ -231,27 +234,6 @@ unsigned long shrink_slab(unsigned long 
 	return ret;
 }
 
-/* Called without lock on whether page is mapped, so answer is unstable */
-static inline int page_mapping_inuse(struct page *page)
-{
-	struct address_space *mapping;
-
-	/* Page is in somebody's page tables. */
-	if (page_mapped(page))
-		return 1;
-
-	/* Be more reluctant to reclaim swapcache than pagecache */
-	if (PageSwapCache(page))
-		return 1;
-
-	mapping = page_mapping(page);
-	if (!mapping)
-		return 0;
-
-	/* File is mmap'd by somebody? */
-	return mapping_mapped(mapping);
-}
-
 static inline int is_page_cache_freeable(struct page *page)
 {
 	return page_count(page) - !!PagePrivate(page) == 2;
@@ -481,7 +463,7 @@ static unsigned long shrink_page_list(st
 
 		referenced = page_referenced(page, 1);
 		/* In active use or really unfreeable?  Activate it. */
-		if (referenced && page_mapping_inuse(page))
+		if (referenced)
 			goto activate_locked;
 
 #ifdef CONFIG_SWAP
@@ -514,8 +496,6 @@ static unsigned long shrink_page_list(st
 		}
 
 		if (PageDirty(page)) {
-			if (referenced)
-				goto keep_locked;
 			if (!may_enter_fs)
 				goto keep_locked;
 			if (!sc->may_writepage)
@@ -595,6 +575,7 @@ keep:
 	if (pagevec_count(&freed_pvec))
 		__pagevec_release_nonlru(&freed_pvec);
 	count_vm_events(PGACTIVATE, pgactivate);
+	sc->activated = pgactivate;
 	return nr_reclaimed;
 }
 
@@ -655,7 +636,7 @@ static unsigned long isolate_lru_pages(u
  * of reclaimed pages
  */
 static unsigned long shrink_inactive_list(unsigned long max_scan,
-				struct zone *zone, struct scan_control *sc)
+			struct zone *zone, struct scan_control *sc, int file)
 {
 	LIST_HEAD(page_list);
 	struct pagevec pvec;
@@ -672,10 +653,17 @@ static unsigned long shrink_inactive_lis
 		unsigned long nr_scan;
 		unsigned long nr_freed;
 
-		nr_taken = isolate_lru_pages(sc->swap_cluster_max,
-					     &zone->inactive_list,
-					     &page_list, &nr_scan);
-		__mod_zone_page_state(zone, NR_INACTIVE, -nr_taken);
+		if (file) {
+			nr_taken = isolate_lru_pages(sc->swap_cluster_max,
+						     &zone->inactive_file_list,
+						     &page_list, &nr_scan);
+			__mod_zone_page_state(zone, NR_INACTIVE_FILE, -nr_taken);
+		} else {
+			nr_taken = isolate_lru_pages(sc->swap_cluster_max,
+						     &zone->inactive_anon_list,
+						     &page_list, &nr_scan);
+			__mod_zone_page_state(zone, NR_INACTIVE_ANON, -nr_taken);
+		}
 		zone->pages_scanned += nr_scan;
 		spin_unlock_irq(&zone->lru_lock);
 
@@ -702,10 +690,21 @@ static unsigned long shrink_inactive_lis
 			VM_BUG_ON(PageLRU(page));
 			SetPageLRU(page);
 			list_del(&page->lru);
-			if (PageActive(page))
-				add_page_to_active_list(zone, page);
-			else
-				add_page_to_inactive_list(zone, page);
+			if (file) {
+			    zone->recent_rotated_file += sc->activated;
+			    zone->recent_scanned_file += nr_scanned;
+			    if (PageActive(page))
+				add_page_to_active_file_list(zone, page);
+			    else
+				add_page_to_inactive_file_list(zone, page);
+			} else {
+			    zone->recent_rotated_anon += sc->activated;
+			    zone->recent_scanned_anon += nr_scanned;
+			    if (PageActive(page))
+				add_page_to_active_anon_list(zone, page);
+			    else
+				add_page_to_inactive_anon_list(zone, page);
+			}
 			if (!pagevec_add(&pvec, page)) {
 				spin_unlock_irq(&zone->lru_lock);
 				__pagevec_release(&pvec);
@@ -736,8 +735,10 @@ static inline void note_zone_scanning_pr
 
 static inline int zone_is_near_oom(struct zone *zone)
 {
-	return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
-				+ zone_page_state(zone, NR_INACTIVE))*3;
+	return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE_ANON)
+				+ zone_page_state(zone, NR_ACTIVE_FILE)
+				+ zone_page_state(zone, NR_INACTIVE_ANON)
+				+ zone_page_state(zone, NR_INACTIVE_FILE))*3;
 }
 
 /*
@@ -758,7 +759,7 @@ static inline int zone_is_near_oom(struc
  * But we had to alter page->flags anyway.
  */
 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
-				struct scan_control *sc, int priority)
+				struct scan_control *sc, int priority, int file)
 {
 	unsigned long pgmoved;
 	int pgdeactivate = 0;
@@ -768,74 +769,28 @@ static void shrink_active_list(unsigned 
 	LIST_HEAD(l_active);	/* Pages to go onto the active_list */
 	struct page *page;
 	struct pagevec pvec;
-	int reclaim_mapped = 0;
-
-	if (sc->may_swap) {
-		long mapped_ratio;
-		long distress;
-		long swap_tendency;
-
-		if (zone_is_near_oom(zone))
-			goto force_reclaim_mapped;
-
-		/*
-		 * `distress' is a measure of how much trouble we're having
-		 * reclaiming pages.  0 -> no problems.  100 -> great trouble.
-		 */
-		distress = 100 >> min(zone->prev_priority, priority);
-
-		/*
-		 * The point of this algorithm is to decide when to start
-		 * reclaiming mapped memory instead of just pagecache.  Work out
-		 * how much memory
-		 * is mapped.
-		 */
-		mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
-				global_page_state(NR_ANON_PAGES)) * 100) /
-					vm_total_pages;
-
-		/*
-		 * Now decide how much we really want to unmap some pages.  The
-		 * mapped ratio is downgraded - just because there's a lot of
-		 * mapped memory doesn't necessarily mean that page reclaim
-		 * isn't succeeding.
-		 *
-		 * The distress ratio is important - we don't want to start
-		 * going oom.
-		 *
-		 * A 100% value of vm_swappiness overrides this algorithm
-		 * altogether.
-		 */
-		swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
-
-		/*
-		 * Now use this metric to decide whether to start moving mapped
-		 * memory onto the inactive list.
-		 */
-		if (swap_tendency >= 100)
-force_reclaim_mapped:
-			reclaim_mapped = 1;
-	}
 
 	lru_add_drain();
 	spin_lock_irq(&zone->lru_lock);
-	pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
-				    &l_hold, &pgscanned);
+	if (file) {
+		pgmoved = isolate_lru_pages(nr_pages, &zone->active_file_list,
+					    &l_hold, &pgscanned);
+		__mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
+	} else {
+		pgmoved = isolate_lru_pages(nr_pages, &zone->active_anon_list,
+					    &l_hold, &pgscanned);
+		__mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
+	}
 	zone->pages_scanned += pgscanned;
-	__mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
 	spin_unlock_irq(&zone->lru_lock);
 
 	while (!list_empty(&l_hold)) {
 		cond_resched();
 		page = lru_to_page(&l_hold);
 		list_del(&page->lru);
-		if (page_mapped(page)) {
-			if (!reclaim_mapped ||
-			    (total_swap_pages == 0 && PageAnon(page)) ||
-			    page_referenced(page, 0)) {
-				list_add(&page->lru, &l_active);
-				continue;
-			}
+		if (page_referenced(page, 0)) {
+			list_add(&page->lru, &l_active);
+			continue;
 		}
 		list_add(&page->lru, &l_inactive);
 	}
@@ -851,10 +806,16 @@ force_reclaim_mapped:
 		VM_BUG_ON(!PageActive(page));
 		ClearPageActive(page);
 
-		list_move(&page->lru, &zone->inactive_list);
+		if (file)
+			list_move(&page->lru, &zone->inactive_file_list);
+		else
+			list_move(&page->lru, &zone->inactive_anon_list);
 		pgmoved++;
 		if (!pagevec_add(&pvec, page)) {
-			__mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
+			if (file)
+				__mod_zone_page_state(zone, NR_INACTIVE_FILE, pgmoved);
+			else
+				__mod_zone_page_state(zone, NR_INACTIVE_ANON, pgmoved);
 			spin_unlock_irq(&zone->lru_lock);
 			pgdeactivate += pgmoved;
 			pgmoved = 0;
@@ -864,7 +825,10 @@ force_reclaim_mapped:
 			spin_lock_irq(&zone->lru_lock);
 		}
 	}
-	__mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
+	if (file)
+		__mod_zone_page_state(zone, NR_INACTIVE_FILE, pgmoved);
+	else
+		__mod_zone_page_state(zone, NR_INACTIVE_ANON, pgmoved);
 	pgdeactivate += pgmoved;
 	if (buffer_heads_over_limit) {
 		spin_unlock_irq(&zone->lru_lock);
@@ -879,17 +843,34 @@ force_reclaim_mapped:
 		VM_BUG_ON(PageLRU(page));
 		SetPageLRU(page);
 		VM_BUG_ON(!PageActive(page));
-		list_move(&page->lru, &zone->active_list);
+		if (file)
+			list_move(&page->lru, &zone->active_file_list);
+		else
+			list_move(&page->lru, &zone->active_anon_list);
 		pgmoved++;
 		if (!pagevec_add(&pvec, page)) {
-			__mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
+			if (file) {
+				__mod_zone_page_state(zone, NR_ACTIVE_FILE, pgmoved);
+				zone->recent_rotated_file += pgmoved;
+			} else {
+				__mod_zone_page_state(zone, NR_ACTIVE_ANON, pgmoved);
+				zone->recent_rotated_anon += pgmoved;
+			}
 			pgmoved = 0;
 			spin_unlock_irq(&zone->lru_lock);
 			__pagevec_release(&pvec);
 			spin_lock_irq(&zone->lru_lock);
 		}
 	}
-	__mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
+	if (file) {
+		__mod_zone_page_state(zone, NR_ACTIVE_FILE, pgmoved);
+		zone->recent_rotated_file += pgmoved;
+		zone->recent_scanned_file += pgscanned;
+	} else {
+		__mod_zone_page_state(zone, NR_ACTIVE_ANON, pgmoved);
+		zone->recent_rotated_anon += pgmoved;
+		zone->recent_scanned_anon += pgscanned;
+	}
 
 	__count_zone_vm_events(PGREFILL, zone, pgscanned);
 	__count_vm_events(PGDEACTIVATE, pgdeactivate);
@@ -899,52 +880,152 @@ force_reclaim_mapped:
 }
 
 /*
+ * The utility of the anon and file memory corresponds to the fraction
+ * of pages that were recently referenced in each category.  Pageout
+ * pressure is distributed according to the size of each set, the fraction
+ * of recently referenced pages (except used-once file pages) and the
+ * swappiness parameter. Note that shrink_zone takes care of size scaling.
+ *
+ * We return the relative pressures as percentages so shrink_zone can
+ * easily use them.
+ */
+static void get_scan_ratio(struct zone *zone, struct scan_control * sc,
+		unsigned long *anon_percent, unsigned long *file_percent)
+{
+	unsigned long anon, file;
+	unsigned long anon_prio, file_prio;
+	unsigned long anon_l, file_l;
+
+	anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
+		zone_page_state(zone, NR_INACTIVE_ANON);
+	file  = zone_page_state(zone, NR_ACTIVE_FILE) +
+		zone_page_state(zone, NR_INACTIVE_FILE);
+
+	/* Keep a floating average of RECENT references. */
+	if (unlikely(zone->recent_rotated_anon > anon / 4)) {
+		spin_lock_irq(&zone->lru_lock);
+		zone->recent_rotated_anon /= 2;
+		zone->recent_scanned_anon /= 2;
+		spin_unlock_irq(&zone->lru_lock);
+	}
+
+	if (unlikely(zone->recent_rotated_file > file / 4)) {
+		spin_lock_irq(&zone->lru_lock);
+		zone->recent_rotated_file /= 2;
+		zone->recent_scanned_file /= 2;
+		spin_unlock_irq(&zone->lru_lock);
+	}
+
+	/*
+	 * With swappiness at 100, anonymous and file have the same priority.
+	 * This scanning priority is essentially the inverse of IO cost.
+	 */
+	anon_prio = sc->swappiness;
+	file_prio = 200 - sc->swappiness;
+
+	/*
+ 	 *                  anon       recent_rotated_anon 
+	 * %anon = 100 * ----------- / ------------------- * IO cost
+	 *               anon + file       rotate_sum
+	 */
+	anon_l = (anon_prio + 1) * (zone->recent_scanned_anon + 1);
+	anon_l /= zone->recent_rotated_anon + 1;
+
+	file_l = (file_prio + 1) * (zone->recent_scanned_file + 1);
+	file_l /= zone->recent_rotated_file + 1;
+
+	/* Normalize to percentages. */
+	*anon_percent = 100 * anon_l / (anon_l + file_l);
+	*file_percent = 100 - *anon_percent;
+}
+
+/*
  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
+ *
+ * Filesystem backed pages live on a separate set of pageout queues from
+ * anonymous, swap and tmpfs pages.  This is done because page cache pages
+ * can usually be evicted and faulted back in more cheaply because of
+ * readahead and better IO clustering.  Also, on many systems swap is
+ * undesirable so we do not want to scan through the anonymous pages.
  */
 static unsigned long shrink_zone(int priority, struct zone *zone,
 				struct scan_control *sc)
 {
-	unsigned long nr_active;
-	unsigned long nr_inactive;
+	unsigned long nr_active_file, nr_inactive_file;
+	unsigned long nr_active_anon, nr_inactive_anon;
+	unsigned long anon_percent, file_percent;
 	unsigned long nr_to_scan;
 	unsigned long nr_reclaimed = 0;
 
 	atomic_inc(&zone->reclaim_in_progress);
 
+	get_scan_ratio(zone, sc, &anon_percent, &file_percent);
 	/*
 	 * Add one to `nr_to_scan' just to make sure that the kernel will
-	 * slowly sift through the active list.
+	 * slowly sift through small lists, too.
 	 */
-	zone->nr_scan_active +=
-		(zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
-	nr_active = zone->nr_scan_active;
-	if (nr_active >= sc->swap_cluster_max)
-		zone->nr_scan_active = 0;
+	zone->nr_scan_active_file +=
+		(zone_page_state(zone, NR_ACTIVE_FILE) >> priority) + 1;
+	nr_active_file = zone->nr_scan_active_file * file_percent / 100;
+	if (nr_active_file >= sc->swap_cluster_max)
+		zone->nr_scan_active_file = 0;
+	else
+		nr_active_file = 0;
+
+	zone->nr_scan_inactive_file +=
+		(zone_page_state(zone, NR_INACTIVE_FILE) >> priority) + 1;
+	nr_inactive_file = zone->nr_scan_inactive_file * file_percent / 100;
+	if (nr_inactive_file >= sc->swap_cluster_max)
+		zone->nr_scan_inactive_file = 0;
+	else
+		nr_inactive_file = 0;
+
+	zone->nr_scan_active_anon +=
+		(zone_page_state(zone, NR_ACTIVE_ANON) >> priority) + 1;
+	nr_active_anon = zone->nr_scan_active_anon * anon_percent / 100;
+	if (nr_active_anon >= sc->swap_cluster_max)
+		zone->nr_scan_active_anon = 0;
 	else
-		nr_active = 0;
+		nr_active_anon = 0;
 
-	zone->nr_scan_inactive +=
-		(zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
-	nr_inactive = zone->nr_scan_inactive;
-	if (nr_inactive >= sc->swap_cluster_max)
-		zone->nr_scan_inactive = 0;
+	zone->nr_scan_inactive_anon +=
+		(zone_page_state(zone, NR_INACTIVE_ANON) >> priority) + 1;
+	nr_inactive_anon = zone->nr_scan_inactive_anon * anon_percent / 100;
+	if (nr_inactive_anon >= sc->swap_cluster_max)
+		zone->nr_scan_inactive_anon = 0;
 	else
-		nr_inactive = 0;
+		nr_inactive_anon = 0;
+
+	while (nr_active_file || nr_inactive_file ||
+				nr_active_anon || nr_inactive_anon) {
+		if (nr_active_file) {
+			nr_to_scan = min(nr_active_file,
+					(unsigned long)sc->swap_cluster_max);
+			nr_active_file -= nr_to_scan;
+			shrink_active_list(nr_to_scan, zone, sc, priority, 1);
+		}
+
+		if (nr_inactive_file) {
+			nr_to_scan = min(nr_inactive_file,
+					(unsigned long)sc->swap_cluster_max);
+			nr_inactive_file -= nr_to_scan;
+			nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
+								sc, 1);
+		}
 
-	while (nr_active || nr_inactive) {
-		if (nr_active) {
-			nr_to_scan = min(nr_active,
+		if (nr_active_anon) {
+			nr_to_scan = min(nr_active_anon,
 					(unsigned long)sc->swap_cluster_max);
-			nr_active -= nr_to_scan;
-			shrink_active_list(nr_to_scan, zone, sc, priority);
+			nr_active_anon -= nr_to_scan;
+			shrink_active_list(nr_to_scan, zone, sc, priority, 0);
 		}
 
-		if (nr_inactive) {
-			nr_to_scan = min(nr_inactive,
+		if (nr_inactive_anon) {
+			nr_to_scan = min(nr_inactive_anon,
 					(unsigned long)sc->swap_cluster_max);
-			nr_inactive -= nr_to_scan;
+			nr_inactive_anon -= nr_to_scan;
 			nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
-								sc);
+								sc, 0);
 		}
 	}
 
@@ -1036,8 +1117,10 @@ unsigned long try_to_free_pages(struct z
 		if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
 			continue;
 
-		lru_pages += zone_page_state(zone, NR_ACTIVE)
-				+ zone_page_state(zone, NR_INACTIVE);
+		lru_pages += zone_page_state(zone, NR_ACTIVE_ANON)
+				+ zone_page_state(zone, NR_ACTIVE_FILE)
+				+ zone_page_state(zone, NR_INACTIVE_ANON)
+				+ zone_page_state(zone, NR_INACTIVE_FILE);
 	}
 
 	for (priority = DEF_PRIORITY; priority >= 0; priority--) {
@@ -1182,8 +1265,10 @@ loop_again:
 		for (i = 0; i <= end_zone; i++) {
 			struct zone *zone = pgdat->node_zones + i;
 
-			lru_pages += zone_page_state(zone, NR_ACTIVE)
-					+ zone_page_state(zone, NR_INACTIVE);
+			lru_pages += zone_page_state(zone, NR_ACTIVE_ANON)
+				   + zone_page_state(zone, NR_ACTIVE_FILE) 
+				   + zone_page_state(zone, NR_INACTIVE_ANON) 
+				   + zone_page_state(zone, NR_INACTIVE_FILE);
 		}
 
 		/*
@@ -1220,8 +1305,10 @@ loop_again:
 			if (zone->all_unreclaimable)
 				continue;
 			if (nr_slab == 0 && zone->pages_scanned >=
-				(zone_page_state(zone, NR_ACTIVE)
-				+ zone_page_state(zone, NR_INACTIVE)) * 6)
+				(zone_page_state(zone, NR_ACTIVE_ANON)
+				+ zone_page_state(zone, NR_ACTIVE_FILE)
+				+ zone_page_state(zone, NR_INACTIVE_ANON)
+				+ zone_page_state(zone, NR_INACTIVE_FILE)) * 6)
 					zone->all_unreclaimable = 1;
 			/*
 			 * If we've done a decent amount of scanning and
@@ -1392,23 +1479,43 @@ static unsigned long shrink_all_zones(un
 
 		/* For pass = 0 we don't shrink the active list */
 		if (pass > 0) {
-			zone->nr_scan_active +=
-				(zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
-			if (zone->nr_scan_active >= nr_pages || pass > 3) {
-				zone->nr_scan_active = 0;
+			zone->nr_scan_active_file +=
+				(zone_page_state(zone, NR_ACTIVE_FILE) >> prio) + 1;
+			if (zone->nr_scan_active_file >= nr_pages || pass > 3) {
+				zone->nr_scan_active_file = 0;
+				nr_to_scan = min(nr_pages,
+					zone_page_state(zone, NR_ACTIVE_FILE));
+				shrink_active_list(nr_to_scan, zone, sc, prio, 1);
+			}
+
+			zone->nr_scan_active_anon +=
+				(zone_page_state(zone, NR_ACTIVE_ANON) >> prio) + 1;
+			if (zone->nr_scan_active_anon >= nr_pages || pass > 3) {
+				zone->nr_scan_active_anon = 0;
 				nr_to_scan = min(nr_pages,
-					zone_page_state(zone, NR_ACTIVE));
-				shrink_active_list(nr_to_scan, zone, sc, prio);
+					zone_page_state(zone, NR_ACTIVE_ANON));
+				shrink_active_list(nr_to_scan, zone, sc, prio, 0);
 			}
 		}
 
-		zone->nr_scan_inactive +=
-			(zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
-		if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
-			zone->nr_scan_inactive = 0;
+		zone->nr_scan_inactive_file +=
+			(zone_page_state(zone, NR_INACTIVE_FILE) >> prio) + 1;
+		if (zone->nr_scan_inactive_file >= nr_pages || pass > 3) {
+			zone->nr_scan_inactive_file = 0;
+			nr_to_scan = min(nr_pages,
+				zone_page_state(zone, NR_INACTIVE_FILE));
+			ret += shrink_inactive_list(nr_to_scan, zone, sc, 1);
+			if (ret >= nr_pages)
+				return ret;
+		}
+
+		zone->nr_scan_inactive_anon +=
+			(zone_page_state(zone, NR_INACTIVE_ANON) >> prio) + 1;
+		if (zone->nr_scan_inactive_anon >= nr_pages || pass > 3) {
+			zone->nr_scan_inactive_anon = 0;
 			nr_to_scan = min(nr_pages,
-				zone_page_state(zone, NR_INACTIVE));
-			ret += shrink_inactive_list(nr_to_scan, zone, sc);
+				zone_page_state(zone, NR_INACTIVE_ANON));
+			ret += shrink_inactive_list(nr_to_scan, zone, sc, 0);
 			if (ret >= nr_pages)
 				return ret;
 		}
@@ -1419,7 +1526,10 @@ static unsigned long shrink_all_zones(un
 
 static unsigned long count_lru_pages(void)
 {
-	return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
+	return global_page_state(NR_ACTIVE_ANON)
+		+ global_page_state(NR_ACTIVE_FILE)
+		+ global_page_state(NR_INACTIVE_ANON)
+		+ global_page_state(NR_INACTIVE_FILE);
 }
 
 /*
--- linux-2.6.21.noarch/mm/vmstat.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/vmstat.c	2007-07-23 11:42:52.000000000 -0400
@@ -477,8 +477,10 @@ const struct seq_operations fragmentatio
 static const char * const vmstat_text[] = {
 	/* Zoned VM counters */
 	"nr_free_pages",
-	"nr_active",
-	"nr_inactive",
+	"nr_active_anon",
+	"nr_inactive_anon",
+	"nr_active_file",
+	"nr_inactive_file",
 	"nr_anon_pages",
 	"nr_mapped",
 	"nr_file_pages",
@@ -554,7 +556,7 @@ static int zoneinfo_show(struct seq_file
 			   "\n        min      %lu"
 			   "\n        low      %lu"
 			   "\n        high     %lu"
-			   "\n        scanned  %lu (a: %lu i: %lu)"
+			   "\n        scanned  %lu (ao: %lu io: %lu af: %lu if: %lu)"
 			   "\n        spanned  %lu"
 			   "\n        present  %lu",
 			   zone_page_state(zone, NR_FREE_PAGES),
@@ -562,7 +564,10 @@ static int zoneinfo_show(struct seq_file
 			   zone->pages_low,
 			   zone->pages_high,
 			   zone->pages_scanned,
-			   zone->nr_scan_active, zone->nr_scan_inactive,
+			   zone->nr_scan_active_anon,
+			   zone->nr_scan_inactive_anon,
+			   zone->nr_scan_active_file,
+			   zone->nr_scan_inactive_file,
 			   zone->spanned_pages,
 			   zone->present_pages);
 
--- linux-2.6.21.noarch/mm/page-writeback.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/page-writeback.c	2007-07-23 11:42:52.000000000 -0400
@@ -151,8 +151,8 @@ static unsigned long determine_dirtyable
 	unsigned long x;
 
 	x = global_page_state(NR_FREE_PAGES)
-		+ global_page_state(NR_INACTIVE)
-		+ global_page_state(NR_ACTIVE);
+		+ global_page_state(NR_INACTIVE_FILE)
+		+ global_page_state(NR_ACTIVE_FILE);
 	x -= highmem_dirtyable_memory(x);
 	return x + 1;	/* Ensure that we never return 0 */
 }
--- linux-2.6.21.noarch/mm/swap.c.vmsplit	2007-07-05 12:06:19.000000000 -0400
+++ linux-2.6.21.noarch/mm/swap.c	2007-07-23 11:42:52.000000000 -0400
@@ -125,7 +125,10 @@ int rotate_reclaimable_page(struct page 
 	zone = page_zone(page);
 	spin_lock_irqsave(&zone->lru_lock, flags);
 	if (PageLRU(page) && !PageActive(page)) {
-		list_move_tail(&page->lru, &zone->inactive_list);
+		if (unlikely(PageSwapCache(page)))
+			list_move_tail(&page->lru, &zone->inactive_anon_list);
+		else
+			list_move_tail(&page->lru, &zone->inactive_file_list);
 		__count_vm_event(PGROTATED);
 	}
 	if (!test_clear_page_writeback(page))
@@ -137,16 +140,32 @@ int rotate_reclaimable_page(struct page 
 /*
  * FIXME: speed this up?
  */
-void fastcall activate_page(struct page *page)
+void fastcall activate_page_file(struct page *page)
 {
 	struct zone *zone = page_zone(page);
 
 	spin_lock_irq(&zone->lru_lock);
 	if (PageLRU(page) && !PageActive(page)) {
-		del_page_from_inactive_list(zone, page);
+		del_page_from_inactive_file_list(zone, page);
 		SetPageActive(page);
-		add_page_to_active_list(zone, page);
+		add_page_to_active_file_list(zone, page);
 		__count_vm_event(PGACTIVATE);
+		zone->recent_rotated_file++;
+	}
+	spin_unlock_irq(&zone->lru_lock);
+}
+
+void fastcall activate_page_anon(struct page *page)
+{
+	struct zone *zone = page_zone(page);
+
+	spin_lock_irq(&zone->lru_lock);
+	if (PageLRU(page) && !PageActive(page)) {
+		del_page_from_inactive_anon_list(zone, page);
+		SetPageActive(page);
+		add_page_to_active_anon_list(zone, page);
+		__count_vm_event(PGACTIVATE);
+		zone->recent_rotated_anon++;
 	}
 	spin_unlock_irq(&zone->lru_lock);
 }
@@ -161,7 +180,10 @@ void fastcall activate_page(struct page 
 void fastcall mark_page_accessed(struct page *page)
 {
 	if (!PageActive(page) && PageReferenced(page) && PageLRU(page)) {
-		activate_page(page);
+		if (page_anon(page))
+			activate_page_anon(page);
+		else
+			activate_page_file(page);
 		ClearPageReferenced(page);
 	} else if (!PageReferenced(page)) {
 		SetPageReferenced(page);
@@ -174,39 +196,67 @@ EXPORT_SYMBOL(mark_page_accessed);
  * lru_cache_add: add a page to the page lists
  * @page: the page to add
  */
-static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
-static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_add_file_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_add_active_file_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_add_anon_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_add_active_anon_pvecs) = { 0, };
+
+void fastcall lru_cache_add_anon(struct page *page)
+{
+	struct pagevec *pvec = &get_cpu_var(lru_add_anon_pvecs);
+
+	page_cache_get(page);
+	if (!pagevec_add(pvec, page))
+		__pagevec_lru_add_anon(pvec);
+	put_cpu_var(lru_add_anon_pvecs);
+}
+
+void fastcall lru_cache_add_file(struct page *page)
+{
+	struct pagevec *pvec = &get_cpu_var(lru_add_file_pvecs);
+
+	page_cache_get(page);
+	if (!pagevec_add(pvec, page))
+		__pagevec_lru_add_file(pvec);
+	put_cpu_var(lru_add_file_pvecs);
+}
 
-void fastcall lru_cache_add(struct page *page)
+void fastcall lru_cache_add_active_anon(struct page *page)
 {
-	struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
+	struct pagevec *pvec = &get_cpu_var(lru_add_active_anon_pvecs);
 
 	page_cache_get(page);
 	if (!pagevec_add(pvec, page))
-		__pagevec_lru_add(pvec);
-	put_cpu_var(lru_add_pvecs);
+		__pagevec_lru_add_active_anon(pvec);
+	put_cpu_var(lru_add_active_anon_pvecs);
 }
 
-void fastcall lru_cache_add_active(struct page *page)
+void fastcall lru_cache_add_active_file(struct page *page)
 {
-	struct pagevec *pvec = &get_cpu_var(lru_add_active_pvecs);
+	struct pagevec *pvec = &get_cpu_var(lru_add_active_file_pvecs);
 
 	page_cache_get(page);
 	if (!pagevec_add(pvec, page))
-		__pagevec_lru_add_active(pvec);
-	put_cpu_var(lru_add_active_pvecs);
+		__pagevec_lru_add_active_file(pvec);
+	put_cpu_var(lru_add_active_file_pvecs);
 }
 
 static void __lru_add_drain(int cpu)
 {
-	struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
+	struct pagevec *pvec = &per_cpu(lru_add_file_pvecs, cpu);
 
 	/* CPU is dead, so no locking needed. */
 	if (pagevec_count(pvec))
-		__pagevec_lru_add(pvec);
-	pvec = &per_cpu(lru_add_active_pvecs, cpu);
+		__pagevec_lru_add_file(pvec);
+	pvec = &per_cpu(lru_add_active_file_pvecs, cpu);
 	if (pagevec_count(pvec))
-		__pagevec_lru_add_active(pvec);
+		__pagevec_lru_add_active_file(pvec);
+	pvec = &per_cpu(lru_add_anon_pvecs, cpu);
+	if (pagevec_count(pvec))
+		__pagevec_lru_add_anon(pvec);
+	pvec = &per_cpu(lru_add_active_anon_pvecs, cpu);
+	if (pagevec_count(pvec))
+		__pagevec_lru_add_active_anon(pvec);
 }
 
 void lru_add_drain(void)
@@ -348,7 +398,32 @@ void __pagevec_release_nonlru(struct pag
  * Add the passed pages to the LRU, then drop the caller's refcount
  * on them.  Reinitialises the caller's pagevec.
  */
-void __pagevec_lru_add(struct pagevec *pvec)
+void __pagevec_lru_add_file(struct pagevec *pvec)
+{
+	int i;
+	struct zone *zone = NULL;
+
+	for (i = 0; i < pagevec_count(pvec); i++) {
+		struct page *page = pvec->pages[i];
+		struct zone *pagezone = page_zone(page);
+
+		if (pagezone != zone) {
+			if (zone)
+				spin_unlock_irq(&zone->lru_lock);
+			zone = pagezone;
+			spin_lock_irq(&zone->lru_lock);
+		}
+		VM_BUG_ON(PageLRU(page));
+		SetPageLRU(page);
+		add_page_to_inactive_file_list(zone, page);
+	}
+	if (zone)
+		spin_unlock_irq(&zone->lru_lock);
+	release_pages(pvec->pages, pvec->nr, pvec->cold);
+	pagevec_reinit(pvec);
+}
+
+void __pagevec_lru_add_active_file(struct pagevec *pvec)
 {
 	int i;
 	struct zone *zone = NULL;
@@ -365,7 +440,9 @@ void __pagevec_lru_add(struct pagevec *p
 		}
 		VM_BUG_ON(PageLRU(page));
 		SetPageLRU(page);
-		add_page_to_inactive_list(zone, page);
+		VM_BUG_ON(PageActive(page));
+		SetPageActive(page);
+		add_page_to_active_file_list(zone, page);
 	}
 	if (zone)
 		spin_unlock_irq(&zone->lru_lock);
@@ -373,9 +450,32 @@ void __pagevec_lru_add(struct pagevec *p
 	pagevec_reinit(pvec);
 }
 
-EXPORT_SYMBOL(__pagevec_lru_add);
+void __pagevec_lru_add_anon(struct pagevec *pvec)
+{
+	int i;
+	struct zone *zone = NULL;
+
+	for (i = 0; i < pagevec_count(pvec); i++) {
+		struct page *page = pvec->pages[i];
+		struct zone *pagezone = page_zone(page);
+
+		if (pagezone != zone) {
+			if (zone)
+				spin_unlock_irq(&zone->lru_lock);
+			zone = pagezone;
+			spin_lock_irq(&zone->lru_lock);
+		}
+		VM_BUG_ON(PageLRU(page));
+		SetPageLRU(page);
+		add_page_to_inactive_anon_list(zone, page);
+	}
+	if (zone)
+		spin_unlock_irq(&zone->lru_lock);
+	release_pages(pvec->pages, pvec->nr, pvec->cold);
+	pagevec_reinit(pvec);
+}
 
-void __pagevec_lru_add_active(struct pagevec *pvec)
+void __pagevec_lru_add_active_anon(struct pagevec *pvec)
 {
 	int i;
 	struct zone *zone = NULL;
@@ -394,7 +494,7 @@ void __pagevec_lru_add_active(struct pag
 		SetPageLRU(page);
 		VM_BUG_ON(PageActive(page));
 		SetPageActive(page);
-		add_page_to_active_list(zone, page);
+		add_page_to_active_anon_list(zone, page);
 	}
 	if (zone)
 		spin_unlock_irq(&zone->lru_lock);
--- linux-2.6.21.noarch/mm/swapfile.c.vmsplit	2007-07-23 12:13:21.000000000 -0400
+++ linux-2.6.21.noarch/mm/swapfile.c	2007-07-23 12:15:20.000000000 -0400
@@ -519,7 +519,7 @@ static void unuse_pte(struct vm_area_str
 	 * Move the page to the active list so it is not
 	 * immediately swapped out again after swapon.
 	 */
-	activate_page(page);
+	activate_page_anon(page);
 }
 
 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
@@ -621,7 +621,7 @@ static int unuse_mm(struct mm_struct *mm
 		 * Activate page so shrink_cache is unlikely to unmap its
 		 * ptes while lock is dropped, so swapoff can make progress.
 		 */
-		activate_page(page);
+		activate_page_anon(page);
 		unlock_page(page);
 		down_read(&mm->mmap_sem);
 		lock_page(page);

[-- Attachment #3: linux-2.6.22-vmsplit-swapfree.patch --]
[-- Type: text/x-patch, Size: 587 bytes --]

--- linux-2.6.22.noarch/mm/vmscan.c.swapfree	2007-07-14 00:36:26.000000000 -0400
+++ linux-2.6.22.noarch/mm/vmscan.c	2007-07-14 00:35:58.000000000 -0400
@@ -892,6 +892,13 @@ static void get_scan_ratio(struct zone *
 	unsigned long anon_prio, file_prio;
 	unsigned long rotate_sum;
 
+	/* If there is no free swap space, do not scan anonymous pages. */
+	if (nr_swap_pages <= 0) {
+		*anon_percent = 0;
+		*file_percent = 100;
+		return;
+	}
+
 	anon  = zone_page_state(zone, NR_ACTIVE_ANON) +
 		zone_page_state(zone, NR_INACTIVE_ANON);
 	file  = zone_page_state(zone, NR_ACTIVE_FILE) +

Re: How can we make page replacement smarter

[Alan Cox]

> Files are different.  File content tends to be grouped
> in large related chunks, both logically in the file and
> on disk.  Generally there is a lot more file data on a
> system than what fits in memory.

Binary paging patterns don't always look like that unfortunately although
I suspect we might want to be weighing towards paging out R/O file mapped
pages earlier simply because they are bigger linear chunks

Re: How can we make page replacement smarter

[Rik van Riel]

Alan Cox wrote:
>> Files are different.  File content tends to be grouped
>> in large related chunks, both logically in the file and
>> on disk.  Generally there is a lot more file data on a
>> system than what fits in memory.
> 
> Binary paging patterns don't always look like that unfortunately although
> I suspect we might want to be weighing towards paging out R/O file mapped
> pages earlier simply because they are bigger linear chunks

A properly implemented use-once algorithm should be able
to filter out the spatial locality of reference pages from
the temporal locality of reference ones, though...

-- 
Politics is the struggle between those who want to make their country
the best in the world, and those who believe it already is.  Each group
calls the other unpatriotic.

Re: How can we make page replacement smarter

[Al Boldi]

Rik van Riel wrote:
> Al Boldi wrote:
> > Good idea, but unless we understand the problems involved, we are bound
> > to repeat it.  So my first question would be:  Why is swap-in so slow?
> >
> > As I have posted in other threads, swap-in of consecutive pages suffers
> > a 2x slowdown wrt swap-out, whereas swap-in of random pages suffers over
> > 6x slowdown.
> >
> > Because it is hard to quantify the expected swap-in speed for random
> > pages, let's first tackle the swap-in of consecutive pages, which should
> > be at least as fast as swap-out.  So again, why is swap-in so slow?
>
> I suspect that this is a locality of reference issue.
>
> Anonymous memory can get jumbled up by repeated free and
> malloc cycles of many smaller objects.  The amount of
> anonymous memory is often smaller than or roughly the same
> size as system memory.

Sounds exactly like the tmpfs problem.

> Locality of refenence to anonymous memory tends to be
> temporal in nature, with the same sets of pages being
> accessed over and over again.
>
> Files are different.  File content tends to be grouped
> in large related chunks, both logically in the file and
> on disk.  Generally there is a lot more file data on a
> system than what fits in memory.
>
> Locality of reference to file data tends to be spatial
> in nature, with one file access leading up to the system
> accessing "nearby" data.  The data is not necessarily
> touched again any time soon.
>
> > Once we understand this problem, we may be able to suggest a smart
> > improvement.
>
> Like the one on http://linux-mm.org/PageoutFailureModes ?

Interesting to see that there are known problems, but it doesn't seem to list 
the resume-from-disk swap-in slowdown.

> I have the LRU lists split and am working on getting SEQ
> replacement implemented for the anonymous pages.
>
> The most recent (untested) patches are attached.

Applied against 2.6.22; the kernel crashes out on boot.


Thanks!

--
Al

Re: swap-prefetch: A smart way to make good use of idle resources (was: updatedb)

[Al Boldi]

Chris Snook wrote:
> Al Boldi wrote:
> > IMHO, what everybody agrees on, is that swap-prefetch has a positive
> > effect in some cases, and nobody can prove an adverse effect (excluding
> > power consumption).  The reason for this positive effect is also crystal
> > clear: It prefetches from swap on idle into free memory, ie: it doesn't
> > force anybody out, and they are the first to be dropped without further
> > swap-out, which sounds really smart.
> >
> > Conclusion:  Either prove swap-prefetch is broken, or get this merged
> > quick.
>
> If you can't prove why it helps and doesn't hurt, then it's a hack, by
> definition.

Ok, slow down: swap-prefetch isn't a hack.  It's a kernel-thread that adds 
swap-prefetch functionality to the kernel.

> With swap prefetch, we're only optimizing the case when the box isn't
> loaded and there's RAM free, but we're not optimizing the case when the
> box is heavily loaded and we need for RAM to be free.

Exactly, swap-prefetch is very specific, and that's why it's so successful:  
It does one thing, and it does that very well.

> I'm inclined to view swap prefetch as a successful scientific experiment,
> and use that data to inform a more reasoned engineering effort.  If we can
> design something intelligent which happens to behave more or less like
> swap prefetch does under the circumstances where swap prefetch helps, and
> does something else smart under the circumstances where swap prefetch
> makes no discernable difference, it'll be a much bigger improvement.

Well, a swapless OS would really be the ultimate, but that's another thread 
entirely (see thread: '[RFC] VM: I have a dream...')

Don't mistake swap-prefetch as trying to additionally fix swap-in slowdown, 
and if it did, then that would be a hack, but it doesn't.

Instead, understand that swap-prefetch is viable even if all swapper issues 
have been solved, because swapping implies pages being swapped in when 
needed, and swap-prefetch smartly uses idle time to do so.

> Because we cannot prove why the existing patch helps, we cannot say what
> impact it will have when things like virtualization and solid state drives
> radically change the coefficients of the equation we have not solved. 
> Providing a sysctl to turn off a misbehaving feature is a poor substitute
> for doing it right the first time, and leaving it off by default will
> ensure that it only gets used by the handful of people who know enough to
> rebuild with the patch anyway.

But we do know why it helps: a proc eats memory, then page-cache, then swaps 
others out, and then dies to free its memory, and now swap-prefetch comes in 
if the system is idle.  Sounds really smart.

While many people may definitely benefit, others may just want to turn it 
off.  No harm done.


Thanks!

--
Al