[PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Rik van Riel]

Hi,

today I released a new version of my VM patch for 2.4.0-test.

This patch should mostly fix streaming IO performance, due
to the following two features:
- drop_behind(), when we do a readahead, move the pages
  'behind' us to the inactive list .. this way we can do
  streaming IO without putting pressure on the working set
- deactivate pages in generic_file_write(), this does
  basically the same ... by moving the pages we write to 
  to the inactive_dirty list, a big download, etc.. doesn't
  impact the working set of the system

I'm particularly interested in the impact of streaming IO on
the performance of the rest of the system with this patch, but
of course also in the performance of the streaming IO itself.

The patch is available from:

	http://www.surriel.com/patches/

	http://www.surriel.com/patches/2.4.0-t8p1-vmpatch2

regards,

Rik
--
"What you're running that piece of shit Gnome?!?!"
       -- Miguel de Icaza, UKUUG 2000

http://www.conectiva.com/		http://www.surriel.com/

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: [PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Arjan van de Ven]

You wrote:

> - drop_behind(), when we do a readahead, move the pages
>   'behind' us to the inactive list .. this way we can do
>   streaming IO without putting pressure on the working set
> - deactivate pages in generic_file_write(), this does
>   basically the same ... by moving the pages we write to 
>   to the inactive_dirty list, a big download, etc.. doesn't
>   impact the working set of the system

> I'm particularly interested in the impact of streaming IO on
> the performance of the rest of the system with this patch, but
> of course also in the performance of the streaming IO itself.

I must say your enhancements have a very positive effect. 
"dbench 48" results:

test7 plain	35   MB/s
test8p1 riel1	27   MB/s
test8p1 riel2	34.2 MB/s  <-- latest patch

so with this patch the VM approaches test7 behavior _really_ close, but 
with the enhancement in interactive performance!

Greetings,
   Arjan van de Ven
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: [PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Aki M Laukkanen]

On Thu, 31 Aug 2000, Rik van Riel wrote:

Hi Rik, 

> today I released a new version of my VM patch for 2.4.0-test.

I seem to have severe troubles with the VM patch although I'm not
so sure it is the culprit. I'm running a t8-p1 kernel with vm2,
sard (for 2.4.0-t5 but applied without faults) + streamfs on SMP.

I get:
VM: reclaim_page_found -2147483647

and then
kernel BUG at page_alloc.c:91!

Oops is hand written so if it looks wrong it's probably a typo.

eax: 0000001f ebx: c1156ba0 ecx: c6ad0000
esi: 0000012f edi: c68a8420 edx: c0208e6c
Stack: c01dc600 c01dc86e 0000005b c1156ba0 0000012f c68a8420
Call backtrace: [<c01dc600>] [<c01dc86e>] [<c013085c>] [<c0130c4e>]
[<c01252ba>]
[<c01277e0>] [<c011b35a>] [<c011f2eb>] [<c011f4be>] [<c010aba7>]
Code: 0f 0b 83 c4 0c 89 f6 d8 2b 05 80 9c 20 c0 69 c0 f1 f0 f0
 
 >>EIP; c012fee1 <__free_pages_ok+41/34c>   <=====
 Trace; c01dc600 <tvecs+2a40/ca5c>
 Trace; c01dc86e <tvecs+2cae/ca5c>
 Trace; c013085c <__free_pages+14/18>
 Trace; c0130c4e <free_page_and_swap_cache+72/78>
 Trace; c01252ba <zap_page_range+186/210>
 Code;  c012fee1 <__free_pages_ok+41/34c>
 00000000 <_EIP>:
 Code;  c012fee1 <__free_pages_ok+41/34c>   <=====
    0:   0f 0b                     ud2a      <=====
 Code;  c012fee3 <__free_pages_ok+43/34c>
    2:   83 c4 0c                  add    $0xc,%esp
 Code;  c012fee6 <__free_pages_ok+46/34c>
    5:   89 f6                     mov    %esi,%esi
 Code;  c012fee8 <__free_pages_ok+48/34c>
    7:   d8 2b                     fsubrs (%ebx)
 Code;  c012feea <__free_pages_ok+4a/34c>
    9:   05 80 9c 20 c0            add    $0xc0209c80,%eax
 Code;  c012feef <__free_pages_ok+4f/34c>
    e:   69 c0 f1 f0 f0 00         imul   $0xf0f0f1,%eax,%eax 

The workload was hdrbench with 30 input and 10 output files
on the streamfs partition. On background I was running sard
and vmstat. Also I managed to crash it when in X
(doing nothing in particular). I would probably think it 
was streamfs if not for the X incident (module was not
even loaded) and sard patch looks pretty safe. Not to
mention that without the patch it has been running solid.

-- 
D.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: [PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Aki M Laukkanen]

On Sun, 3 Sep 2000, Aki M Laukkanen wrote:
> I seem to have severe troubles with the VM patch although I'm not
> so sure it is the culprit. I'm running a t8-p1 kernel with vm2,
> sard (for 2.4.0-t5 but applied without faults) + streamfs on SMP.

Ok, it seems partly a false alert. I tried combinations streamfs+vmpatch 
and sard+streamfs. Both worked just fine so the combination of vmpatch
and sard patch seemed to trigger the problems.

http://www.cs.helsinki.fi/u/amlaukka/streamfs/hdrbench30x10-vmpatch2.png

It seems the drop-behind stuff is still not very tuned. Compare with 
hdrbench30x10.png and you'll see the difference. Infact hdrbench aborted
the test because of buffer overflows.

-- 
D.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: [PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Christoph Rohland]

Hi Rik,

Rik van Riel <riel@conectiva.com.br> writes:
> today I released a new version of my VM patch for 2.4.0-test.

Are you aware that this breaks shm swapping?

Greetings
		Christoph
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: [PATCH *] VM patch w/ drop behind for 2.4.0-test8-pre1

[Rik van Riel]

On 5 Sep 2000, Christoph Rohland wrote:
> Rik van Riel <riel@conectiva.com.br> writes:
> > today I released a new version of my VM patch for 2.4.0-test.
> 
> Are you aware that this breaks shm swapping?

Nope. Nobody told me yet. I'll work on a new (working)
version today...

regards,

Rik
--
"What you're running that piece of shit Gnome?!?!"
       -- Miguel de Icaza, UKUUG 2000

http://www.conectiva.com/		http://www.surriel.com/

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/