[2.4.20-ck7] good compressed caching experience

[2.4.20-ck7] good compressed caching experience

[Kimmo Sundqvist]

Hello

I just decided to tell everyone that I've been able to run 2.4.20-ck7 with 
compressed caching enabled in my little brother's Pentium 133MHz, for hours, 
doing stress testing, compiling kernels and using the Internet under X.

I had pre-empt enabled.  Compressed swap worked also.  I used 4kB pages 
without compressed swap, and 8kB with it.

This was with Con's ck7pre versions released on 24th and 25th of May.

Now running 2.4.20-ck7pre with compressed cache in a dual CPU machine with SMP 
disabled (compressed caching and SMP support are still mutually exclusive), 
1GB of RAM but "mem=128M" for testing purposes.  Been stable for 6 hours now, 
and done even some stress testing.  Try 128 instances of burnBX with 1MB 
each, like "for ((A=128;A--;A<1)) do burnBX J & done".  A nice brute force or 
"if you don't behave I'll push all my buttons" method :)

Wondering if Pentium 133MHz (64MB RAM) is fast enough to benefit from 
compressed caching.  I know there's a limit, depending on the speed of the 
CPU and the speed of the swap partition (doing random accesses), which 
determines if compressed caching is beneficial or not.

This machine has a Seagate Barracuda V 80GB, which does sequential reads at 
40MB/s.  I could drive this into trashing, then type "sar -B 1 1000" and see 
how the swap is doing.  Now, compressed caching brings me benefit if, and 
only if, it can compress and decompress pages faster than that in this CPU, 
which it sure does, since this is a Pentium III 933MHz, but I'm not sure 
about the little brother's Pentium 133MHz.  It has a 4GB Seagate that does 
6MB/s sequentially.  Did I figure it out correctly?  Of course swapping to a 
partition gets slower as the swap usage increases.  Longer seeks and the 
like.

Just a warning... both systems have only ReiserFS partitions.  Other FSes 
might still get hurt.

-Kimmo Sundqvist

Re: [2.4.20-ck7] good compressed caching experience

[Con Kolivas]

On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:
> Hello

Hi Kimmo

> I just decided to tell everyone that I've been able to run 2.4.20-ck7 with
> compressed caching enabled in my little brother's Pentium 133MHz, for
> hours, doing stress testing, compiling kernels and using the Internet under
> X.
>
> I had pre-empt enabled.  Compressed swap worked also.  I used 4kB pages
> without compressed swap, and 8kB with it.
>
> This was with Con's ck7pre versions released on 24th and 25th of May.
>
> Now running 2.4.20-ck7pre with compressed cache in a dual CPU machine with
> SMP disabled (compressed caching and SMP support are still mutually
> exclusive), 1GB of RAM but "mem=128M" for testing purposes.  Been stable
> for 6 hours now, and done even some stress testing.  Try 128 instances of
> burnBX with 1MB each, like "for ((A=128;A--;A<1)) do burnBX J & done".  A
> nice brute force or "if you don't behave I'll push all my buttons" method
> :)
>
> Wondering if Pentium 133MHz (64MB RAM) is fast enough to benefit from
> compressed caching.  I know there's a limit, depending on the speed of the
> CPU and the speed of the swap partition (doing random accesses), which
> determines if compressed caching is beneficial or not.
>
> This machine has a Seagate Barracuda V 80GB, which does sequential reads at
> 40MB/s.  I could drive this into trashing, then type "sar -B 1 1000" and
> see how the swap is doing.  Now, compressed caching brings me benefit if,
> and only if, it can compress and decompress pages faster than that in this
> CPU, which it sure does, since this is a Pentium III 933MHz, but I'm not
> sure about the little brother's Pentium 133MHz.  It has a 4GB Seagate that
> does 6MB/s sequentially.  Did I figure it out correctly?  Of course
> swapping to a partition gets slower as the swap usage increases.  Longer
> seeks and the like.

What you describe has been my experience with cc as well. I haven't had any 
crashes or unusual problems with it since removing the AA vm changes as well 
- it seemed to be the combination that caused hiccups on extreme testing. 
>From what I can see, no matter how slow your cpu you will still get benefit 
from cc as the hard drives on those machines are proportionately slower as 
well. The one limitation of cc is that it does require _some_ ram to actually 
store swap pages in, and it seems that you need more than 32Mb ram to start 
deriving benefit.

One minor thing, though - my vm hacks make compressed caching work much less 
than it normally does as they try to avoid swapping quite aggressively. It is 
when the vm attempts to start swapping that cc looks to see if it should take 
pages into compressed cache instead.

I've cc'ed the actual developer of cc as he has indicated that he is actively 
working on compressed caching again.

> Just a warning... both systems have only ReiserFS partitions.  Other FSes
> might still get hurt.

This is definitely the case! If you try out compressed caching with ck7 please 
do not enable preempt if you are using ext2/3 or vfat.

Cheers,
Con

Re: [2.4.20-ck7] good compressed caching experience

[Rodrigo Souza de Castro]

Hi Kimmo!

On Mon, May 26, 2003 at 09:50:03PM +0300, Kimmo Sundqvist wrote:
> I just decided to tell everyone that I've been able to run
> 2.4.20-ck7 with compressed caching enabled in my little brother's
> Pentium 133MHz, for hours, doing stress testing, compiling kernels
> and using the Internet under X.

Great, glad to know that. 

> I had pre-empt enabled.  Compressed swap worked also.  I used 4kB
> pages without compressed swap, and 8kB with it.

Compressed cache patch isn't safe to be used with preempt, at least
not with the version you tried (0.24pre5). There are some bug reports
that I was able to check myself about fs corruptions (due to problems
in locking control). The latest code seems to fix these corruptions,
given some tests some volunteers and I performed. I ported this code
to 2.4.20, but I still couldn't figure out why it isn't as stable as
in 2.4.18 version, that is why it remains unreleased as a patch
(although it is in the CVS server).

Double page size (i.e, 8K pages) is useful when you don't have a good
compression ratio. If your data compress to more than 50% of its
original size, a compressed cache with 4K isn't supposed to provide
major performance gains.

> This was with Con's ck7pre versions released on 24th and 25th of
> May.
> 
> Now running 2.4.20-ck7pre with compressed cache in a dual CPU
> machine with SMP disabled (compressed caching and SMP support are
> still mutually exclusive),

Until I make it work safely on SMP systems, I want to make it
exclusive in the config.in.

> 1GB of RAM but "mem=128M" for testing purposes.  Been stable for 6
> hours now, and done even some stress testing.  Try 128 instances of
> burnBX with 1MB each, like "for ((A=128;A--;A<1)) do burnBX J &
> done".  A nice brute force or "if you don't behave I'll push all my
> buttons" method :)

:-)

> Wondering if Pentium 133MHz (64MB RAM) is fast enough to benefit
> from compressed caching.  I know there's a limit, depending on the
> speed of the CPU and the speed of the swap partition (doing random
> accesses), which determines if compressed caching is beneficial or
> not.

That's right. Faster CPUs have a greater tendency to benefit from
compressed caching, since the gap between them and disks is
larger. This gap is the basic principle that makes compressed caching
interesting today.

> This machine has a Seagate Barracuda V 80GB, which does sequential
> reads at 40MB/s.  I could drive this into trashing, then type "sar
> -B 1 1000" and see how the swap is doing.  Now, compressed caching
> brings me benefit if, and only if, it can compress and decompress
> pages faster than that in this CPU, which it sure does, since this
> is a Pentium III 933MHz, but I'm not sure about the little brother's
> Pentium 133MHz.  It has a 4GB Seagate that does 6MB/s sequentially.
> Did I figure it out correctly?  Of course swapping to a partition
> gets slower as the swap usage increases.  Longer seeks and the like.

Yes, you figured it out correctly. I don't know the lower limit for a
CPU to benefit from compressed caching. However I guess that, since we
already had a gap between disk and CPU with a Pentium 133 MHz, it will
probably have some advantage using it, even if not as much as with a
faster CPU. The current code was written and tested on a Pentium III 1
GHz CPU, so there may be room for improvements for such systems, as
your Pentium 133 MHz.

I would like to know your impressions about compressed caching on your
brother's system.

> Just a warning... both systems have only ReiserFS partitions.  Other
> FSes might still get hurt.

True. I don't know how you don't have your filesystem corrupted
though. I suggest you to disable preempt if you want to keep testing
this compressed cache code in order to avoid possible problems (and
you getting mad at me :-).

Thanks for your feedback.

Regards,
-- 
Rodrigo

Re: [2.4.20-ck7] good compressed caching experience

[Rodrigo Souza de Castro]

Hi Con,

On Tue, May 27, 2003 at 07:11:34AM +1000, Con Kolivas wrote:
> On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:
> > I just decided to tell everyone that I've been able to run 2.4.20-ck7 with
> > compressed caching enabled in my little brother's Pentium 133MHz, for
> > hours, doing stress testing, compiling kernels and using the Internet under
> > X.
> >
> > I had pre-empt enabled.  Compressed swap worked also.  I used 4kB pages
> > without compressed swap, and 8kB with it.
[snip]
> > exclusive), 1GB of RAM but "mem=128M" for testing purposes.  Been stable
> > for 6 hours now, and done even some stress testing.  Try 128 instances of
> > burnBX with 1MB each, like "for ((A=128;A--;A<1)) do burnBX J & done".  A
> > nice brute force or "if you don't behave I'll push all my buttons" method
> > :)
> >
> > Wondering if Pentium 133MHz (64MB RAM) is fast enough to benefit from
> > compressed caching.  I know there's a limit, depending on the speed of the
> > CPU and the speed of the swap partition (doing random accesses), which
> > determines if compressed caching is beneficial or not.
[snip]
> What you describe has been my experience with cc as well. I haven't
> had any crashes or unusual problems with it since removing the AA vm
> changes as well - it seemed to be the combination that caused
> hiccups on extreme testing.

Something to be figured out yet. It is a pity we couldn't work harder
on these problems.

> >From what I can see, no matter how slow your cpu you will still get
> benefit from cc as the hard drives on those machines are
> proportionately slower as well.

I guess that, in the past, the gap was smaller, but still there.

> The one limitation of cc is that it does require _some_ ram to
> actually store swap pages in, and it seems that you need more than
> 32Mb ram to start deriving benefit.

I am not sure. It requires some ram, but it is only a few pages to
avoid failed page allocations when it first needs to allocate pages. I
have a report of running CC on a laptop with 8M RAM, that showed to be
a little more responsive. I don't have a scientific results showing
that is better on system with a lower amount of memory, but this
feedback seems positive.

> One minor thing, though - my vm hacks make compressed caching work
> much less than it normally does as they try to avoid swapping quite
> aggressively. It is when the vm attempts to start swapping that cc
> looks to see if it should take pages into compressed cache instead.

What exactly are your VM hacks concerning CC? The default behaviour is
to compress pages only when the VM starts freeing pages, not in a
compress-ahead fashion, pretty much what I think you said above.

> I've cc'ed the actual developer of cc as he has indicated that he is
> actively working on compressed caching again.

At a slower pace, but finally back to CC development. :-)

> > Just a warning... both systems have only ReiserFS partitions.
> > Other FSes might still get hurt.
> 
> This is definitely the case! If you try out compressed caching with
> ck7 please do not enable preempt if you are using ext2/3 or vfat.

As told in a previous email, I wouldn't enable preempt in any case
with this code version.

Regards,
-- 
Rodrigo

Re: [2.4.20-ck7] good compressed caching experience

[Con Kolivas]

On Tue, 27 May 2003 11:41, Rodrigo Souza de Castro wrote:
> Hi Con,
>
> On Tue, May 27, 2003 at 07:11:34AM +1000, Con Kolivas wrote:
> > On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:
> > > I just decided to tell everyone that I've been able to run 2.4.20-ck7
> > > with compressed caching enabled in my little brother's Pentium 133MHz,
> > > for hours, doing stress testing, compiling kernels and using the
> > > Internet under X.
> > >
> > > I had pre-empt enabled.  Compressed swap worked also.  I used 4kB pages
> > > without compressed swap, and 8kB with it.
>
> [snip]
>
> > > exclusive), 1GB of RAM but "mem=128M" for testing purposes.  Been
> > > stable for 6 hours now, and done even some stress testing.  Try 128
> > > instances of burnBX with 1MB each, like "for ((A=128;A--;A<1)) do
> > > burnBX J & done".  A nice brute force or "if you don't behave I'll push
> > > all my buttons" method
> > >
> > > :)
> > >
> > > Wondering if Pentium 133MHz (64MB RAM) is fast enough to benefit from
> > > compressed caching.  I know there's a limit, depending on the speed of
> > > the CPU and the speed of the swap partition (doing random accesses),
> > > which determines if compressed caching is beneficial or not.
>
> [snip]
>
> > What you describe has been my experience with cc as well. I haven't
> > had any crashes or unusual problems with it since removing the AA vm
> > changes as well - it seemed to be the combination that caused
> > hiccups on extreme testing.
>
> Something to be figured out yet. It is a pity we couldn't work harder
> on these problems.

I think getting the stability necessary on the main tree was far more 
important so I don't regret this.

>
> > >From what I can see, no matter how slow your cpu you will still get
> >
> > benefit from cc as the hard drives on those machines are
> > proportionately slower as well.
>
> I guess that, in the past, the gap was smaller, but still there.
>
> > The one limitation of cc is that it does require _some_ ram to
> > actually store swap pages in, and it seems that you need more than
> > 32Mb ram to start deriving benefit.
>
> I am not sure. It requires some ram, but it is only a few pages to
> avoid failed page allocations when it first needs to allocate pages. I
> have a report of running CC on a laptop with 8M RAM, that showed to be
> a little more responsive. I don't have a scientific results showing
> that is better on system with a lower amount of memory, but this
> feedback seems positive.
>
> > One minor thing, though - my vm hacks make compressed caching work
> > much less than it normally does as they try to avoid swapping quite
> > aggressively. It is when the vm attempts to start swapping that cc
> > looks to see if it should take pages into compressed cache instead.
>
> What exactly are your VM hacks concerning CC? The default behaviour is
> to compress pages only when the VM starts freeing pages, not in a
> compress-ahead fashion, pretty much what I think you said above.

Ok, I didn't look at your code, but that would make sense too because with my 
hacks it will start even lower priority than the default VM freeing less 
pages at a time unless the memory pressure gets high.

>
> > I've cc'ed the actual developer of cc as he has indicated that he is
> > actively working on compressed caching again.
>
> At a slower pace, but finally back to CC development. :-)

Excellent. Good to have you back.
>
> > > Just a warning... both systems have only ReiserFS partitions.
> > > Other FSes might still get hurt.
> >
> > This is definitely the case! If you try out compressed caching with
> > ck7 please do not enable preempt if you are using ext2/3 or vfat.
>
> As told in a previous email, I wouldn't enable preempt in any case
> with this code version.

-ck has always come with a warning saying as much. I hope to integrate more of 
your code when you're happy with it, so this problem can be resolved.

Cheers,
Con

Re: [2.4.20-ck7] good compressed caching experience

[Kimmo Sundqvist]

On Tuesday 27 May 2003 00:11, Con Kolivas wrote:
> On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:

> > Just a warning... both systems have only ReiserFS partitions.  Other FSes
> > might still get hurt.

> This is definitely the case! If you try out compressed caching with ck7
> please do not enable preempt if you are using ext2/3 or vfat.

Is this a problem in ext2/3, pre-empt implementation, compressed caching or 
kernel in general?

I think I can still choose between compression methods, or can I?  Which one 
of them, on average, is the least CPU-intensive, and which one gives the best 
compression ratio?  I am also at loss how to interpret the percentages in 
"cat /proc/comp_cache_stat".

For M$ Windows there was once a program called MagnaRAM97 that had a similar 
idea, but I don't understand how it could report 2 to 3-fold compression 
ratios.  It always spontaneously rebooted the Pentium 133MHz after some 
hours, so I uninstalled it.

Just take your time, but will we see a pre-empt safe (or better yet SMP safe) 
version coming out anytime soon?

Compiling another 2.4.20-ck7 with 8kB pages and swap compression in the 
background.  I have now "mem=896M" to avoid the highmem boundary, even if it 
wasn't necessary.  Someone said somewhere that a 1GB system is faster without 
highmem support, so I haven't compiled it in for a while.

-Kimmo S.

Re: [2.4.20-ck7] good compressed caching experience

[Con Kolivas]

On Wed, 28 May 2003 00:13, Kimmo Sundqvist wrote:
> On Tuesday 27 May 2003 00:11, Con Kolivas wrote:
> > On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:
> > > Just a warning... both systems have only ReiserFS partitions.  Other
> > > FSes might still get hurt.
> >
> > This is definitely the case! If you try out compressed caching with ck7
> > please do not enable preempt if you are using ext2/3 or vfat.
>
> Is this a problem in ext2/3, pre-empt implementation, compressed caching or
> kernel in general?

The compressed caching code isn't yet properly preempt aware. The filesystems 
that are safe are just lucky.

>
> I think I can still choose between compression methods, or can I?  Which
> one of them, on average, is the least CPU-intensive, and which one gives
> the best compression ratio?  I am also at loss how to interpret the
> percentages in "cat /proc/comp_cache_stat".

The default in -ck7 is lzo which is the best compression and pretty fast. The 
full details only RCastro can tell us.

>
> For M$ Windows there was once a program called MagnaRAM97 that had a
> similar idea, but I don't understand how it could report 2 to 3-fold
> compression ratios.  It always spontaneously rebooted the Pentium 133MHz
> after some hours, so I uninstalled it.
>
> Just take your time, but will we see a pre-empt safe (or better yet SMP
> safe) version coming out anytime soon?
>
> Compiling another 2.4.20-ck7 with 8kB pages and swap compression in the
> background.  I have now "mem=896M" to avoid the highmem boundary, even if
> it wasn't necessary.  Someone said somewhere that a 1GB system is faster
> without highmem support, so I haven't compiled it in for a while.

Definitely with the default VM as the basis in 2.4 (as opposed to AA or RMAP) 
you are much better off without High Mem enabled if you only have 1Gb.

Con

Re: [2.4.20-ck7] good compressed caching experience

[Rodrigo Souza de Castro]

Hey Kimmo,

On Tue, May 27, 2003 at 05:13:48PM +0300, Kimmo Sundqvist wrote:
> On Tuesday 27 May 2003 00:11, Con Kolivas wrote:
> > On Tue, 27 May 2003 04:50, Kimmo Sundqvist wrote:
> > > Just a warning... both systems have only ReiserFS partitions.  Other FSes
> > > might still get hurt.
> 
> > This is definitely the case! If you try out compressed caching with ck7
> > please do not enable preempt if you are using ext2/3 or vfat.
> 
> Is this a problem in ext2/3, pre-empt implementation, compressed caching or 
> kernel in general?

It is a compressed caching problem. As Con said, FSs that are safe are
just lucky.

> I think I can still choose between compression methods, or can I?

You can, just use the kernel parameter "compalg=" kernel parameter. 

compalg=0 (WKdm)
compalg=1 (WK4x4)
compalg=2 (LZO, default)

> Which one of them, on average, is the least CPU-intensive, and which
> one gives the best compression ratio?  I am also at loss how to
> interpret the percentages in "cat /proc/comp_cache_stat".

LZO is the most CPU-intensive and gives the best compression
ratio. Even if spending more CPU cycles to compress than WKdm or
WK4x4, it showed to have the best performance given its compression
ratio.

> For M$ Windows there was once a program called MagnaRAM97 that had a
> similar idea, but I don't understand how it could report 2 to 3-fold
> compression ratios.  It always spontaneously rebooted the Pentium
> 133MHz after some hours, so I uninstalled it.

I am not sure, but I guess the purpose of MagnaRAM and this patch are
different, as well as the approach. This patch intends to reduce the
number of disk accesses using compression and extending effective
memory size, therefore improving general performance. We use an
adaptive cache size to avoid overhead when compression isn't/wouldn't
help.

> Just take your time, but will we see a pre-empt safe (or better yet
> SMP safe) version coming out anytime soon?

Yes, you will, they are the top items on my todo list.

> Compiling another 2.4.20-ck7 with 8kB pages and swap compression in
> the background.  I have now "mem=896M" to avoid the highmem
> boundary, even if it wasn't necessary.  Someone said somewhere that
> a 1GB system is faster without highmem support, so I haven't
> compiled it in for a while.

Best regards,
-- 
Rodrigo