* Writeout in recent kernels/VMs poor compared to last -ac
@ 2002-01-12 15:17 Adam Kropelin
2002-01-12 15:29 ` Adam Kropelin
` (3 more replies)
0 siblings, 4 replies; 7+ messages in thread
From: Adam Kropelin @ 2002-01-12 15:17 UTC (permalink / raw)
To: linux-kernel
I recently began regularly transferring large (600 MB+) files to my
Linux-based fileserver and have noticed what I would characterize as poor
writeout behavior under this load. I've done a bit of comparison testing
which may help reveal the problem better.
Disclaimer: I did not choose this test because it is scientific or stresses
the system in any particular way. I use it because it's an operation which I
perform frequently and so the faster it goes, the broader I smile. This may
be an entirely inappropriate load to optimize for, and if so I understand.
Test consisted of FTPing in (Linux was the destination) a 650 MB file and
timing the transfer (from the client's perspective) which capturing "vmstat
1" output on the server. Server hardware is Dual PPro 200, 160 MB RAM, 512
MB swap, destination filesystem is VFAT[0] on cpqarray partition on a RAID5
logical drive. Root and swap are not on the cpqarray.
Here are my results (average of 2 runs, each from a clean reboot with all
unneeded services disabled):
2.4.17: 6:52
2.4.17-rmap11a: 6:53
2.4.18-pre2aa2: 7:10
2.4.13-ac7: 5:30 (!)
2.4.17 with cpqarray driver update from -ac: 6:30
The last test was just to see if -ac's better performance had anything to do
with the driver update. Apparently it had little or nothing to do with it.
The vmstat output is very revealing. All tests except for -ac show a strong
oscillation on the blocks out (this is a representative sample from stock
2.4.17 but the other recent kernels show essentially the same behavior):
procs memory swap io system
cpu
r b w swpd free buff cache si so bi bo in cs us sy
id
0 1 1 0 4408 2712 118756 0 0 0 0 109 16 0 0
100
0 1 1 0 4436 2716 118724 0 0 1 3913 287 13 0 8
92
0 1 0 0 4436 2716 118724 0 0 0 0 118 11 0 0
100
0 0 0 0 4352 2724 118816 0 0 8 0 3639 203 2 28
70
0 0 0 0 4420 2736 118752 0 0 11 0 4530 259 0 33
67
0 0 0 0 4348 2744 118816 0 0 10 0 4273 245 0 42
58
1 0 0 0 4376 2756 118772 0 0 11 0 4551 246 1 39
60
0 1 1 0 4364 2760 118724 0 0 4 6730 1710 93 1 19
80
0 1 1 0 4364 2760 118724 0 0 0 0 108 9 0 0
100
0 1 1 0 4364 2760 118724 0 0 0 3667 117 9 0 2
98
0 1 1 0 4364 2760 118724 0 0 0 0 125 8 0 0
100
1 0 1 0 4364 2760 118724 0 0 0 3819 124 10 0 2
98
0 1 1 0 4372 2760 118704 0 0 0 2055 195 10 0 5
95
0 1 1 0 4372 2760 118704 0 0 0 0 120 6 0 0
100
0 1 1 0 4408 2760 118668 0 0 0 2415 203 16 0 4
96
0 1 1 0 4472 2760 118608 0 0 0 4321 280 18 1 6
93
0 2 1 0 4468 2760 118608 0 0 0 0 112 8 0 0
100
0 1 1 0 4352 2760 118724 0 0 1 3232 227 9 0 10
90
0 1 1 0 4352 2760 118724 0 0 0 0 108 8 0 0
100
0 0 0 0 4440 2768 118696 0 0 7 0 3233 179 1 25
74
1 0 0 0 4448 2776 118680 0 0 11 0 4417 253 0 36
63
-ac, on the other hand, is very smooth (still a noticeable oscillation, but
far more consistent):
procs memory swap io system
cpu
r b w swpd free buff cache si so bi bo in cs us sy
id
0 1 0 0 3064 3900 121580 0 0 4 2051 1714 174 0 16
83
0 1 0 0 3064 3904 121576 0 0 3 1572 1331 133 0 10
90
0 1 0 0 3064 3904 121576 0 0 4 2048 1727 163 0 14
86
0 1 0 0 3064 3908 121572 0 0 4 2048 1720 160 0 17
82
0 1 0 0 3064 3912 121568 0 0 3 1536 1312 124 0 12
88
0 1 0 0 3064 3916 121564 0 0 4 2048 1727 168 0 16
84
0 1 1 0 3064 3920 121560 0 0 4 2080 1720 158 0 16
84
0 1 0 0 3064 3924 121556 0 0 3 1536 1314 119 0 12
87
0 1 0 0 3064 3928 121552 0 0 4 2048 1719 160 1 13
86
0 1 1 0 3064 3932 121548 0 0 4 2048 1726 159 1 15
84
0 1 0 0 3064 3936 121544 0 0 3 1536 1327 129 1 12
86
0 1 0 0 3064 3940 121540 0 0 4 2048 1714 159 0 18
82
1 0 1 0 3064 3944 121536 0 0 3 1920 1604 151 0 16
83
0 1 0 0 3064 3944 121536 0 0 4 1696 1429 130 1 13
86
0 1 0 0 3064 3948 121532 0 0 4 2048 1726 159 1 17
82
0 1 0 0 3064 3952 121528 0 0 3 1536 1324 121 0 11
88
0 1 0 0 3064 3956 121524 0 0 4 2051 1722 164 1 15
83
0 0 1 0 3064 3960 121520 0 0 3 1728 1586 139 1 13
86
0 1 0 0 3064 3964 121516 0 0 4 1856 1455 129 1 12
86
0 1 0 0 3064 3968 121512 0 0 4 2052 1717 162 0 17
83
0 1 0 0 3064 3972 121508 0 0 3 1568 1327 131 0 12
88
Full vmstat output for all runs is available by request.
Any ideas? I'm willing to do further testing if it is of use to anyone. I
rather like the ~15% boost I get from -ac, and I feel much more comfortable
seeing consistent behavior.
--Adam
[0] Early on I experienced some flukes with the cpqarray driver (bad RAM, I
think) so I use VFAT in case Linux suddenly doesn't want to access the array
anymore I will have a chance of booting $OTHER_OS and recovering my data.
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 15:17 Writeout in recent kernels/VMs poor compared to last -ac Adam Kropelin
@ 2002-01-12 15:29 ` Adam Kropelin
2002-01-12 16:03 ` Rik van Riel
` (2 subsequent siblings)
3 siblings, 0 replies; 7+ messages in thread
From: Adam Kropelin @ 2002-01-12 15:29 UTC (permalink / raw)
To: linux-kernel
Arg... Sorry for the wrapped vmstat output. This should be better...
--Adam
procs memory swap io system cpu
r b w swpd free buff cache si so bi bo in cs us sy id
0 1 1 0 4408 2712 118756 0 0 0 0 109 16 0 0 100
0 1 1 0 4436 2716 118724 0 0 1 3913 287 13 0 8 92
0 1 0 0 4436 2716 118724 0 0 0 0 118 11 0 0 100
0 0 0 0 4352 2724 118816 0 0 8 0 3639 203 2 28 70
0 0 0 0 4420 2736 118752 0 0 11 0 4530 259 0 33 67
0 0 0 0 4348 2744 118816 0 0 10 0 4273 245 0 42 58
1 0 0 0 4376 2756 118772 0 0 11 0 4551 246 1 39 60
0 1 1 0 4364 2760 118724 0 0 4 6730 1710 93 1 19 80
0 1 1 0 4364 2760 118724 0 0 0 0 108 9 0 0 100
0 1 1 0 4364 2760 118724 0 0 0 3667 117 9 0 2 98
0 1 1 0 4364 2760 118724 0 0 0 0 125 8 0 0 100
1 0 1 0 4364 2760 118724 0 0 0 3819 124 10 0 2 98
0 1 1 0 4372 2760 118704 0 0 0 2055 195 10 0 5 95
0 1 1 0 4372 2760 118704 0 0 0 0 120 6 0 0 100
0 1 1 0 4408 2760 118668 0 0 0 2415 203 16 0 4 96
0 1 1 0 4472 2760 118608 0 0 0 4321 280 18 1 6 93
0 2 1 0 4468 2760 118608 0 0 0 0 112 8 0 0 100
0 1 1 0 4352 2760 118724 0 0 1 3232 227 9 0 10 90
0 1 1 0 4352 2760 118724 0 0 0 0 108 8 0 0 100
0 0 0 0 4440 2768 118696 0 0 7 0 3233 179 1 25 74
1 0 0 0 4448 2776 118680 0 0 11 0 4417 253 0 36 63
-ac, on the other hand, is very smooth (still a noticeable oscillation, but
far more consistent):
procs memory swap io system cpu
r b w swpd free buff cache si so bi bo in cs us sy id
0 1 0 0 3064 3900 121580 0 0 4 2051 1714 174 0 16 83
0 1 0 0 3064 3904 121576 0 0 3 1572 1331 133 0 10 90
0 1 0 0 3064 3904 121576 0 0 4 2048 1727 163 0 14 86
0 1 0 0 3064 3908 121572 0 0 4 2048 1720 160 0 17 82
0 1 0 0 3064 3912 121568 0 0 3 1536 1312 124 0 12 88
0 1 0 0 3064 3916 121564 0 0 4 2048 1727 168 0 16 84
0 1 1 0 3064 3920 121560 0 0 4 2080 1720 158 0 16 84
0 1 0 0 3064 3924 121556 0 0 3 1536 1314 119 0 12 87
0 1 0 0 3064 3928 121552 0 0 4 2048 1719 160 1 13 86
0 1 1 0 3064 3932 121548 0 0 4 2048 1726 159 1 15 84
0 1 0 0 3064 3936 121544 0 0 3 1536 1327 129 1 12 86
0 1 0 0 3064 3940 121540 0 0 4 2048 1714 159 0 18 82
1 0 1 0 3064 3944 121536 0 0 3 1920 1604 151 0 16 83
0 1 0 0 3064 3944 121536 0 0 4 1696 1429 130 1 13 86
0 1 0 0 3064 3948 121532 0 0 4 2048 1726 159 1 17 82
0 1 0 0 3064 3952 121528 0 0 3 1536 1324 121 0 11 88
0 1 0 0 3064 3956 121524 0 0 4 2051 1722 164 1 15 83
0 0 1 0 3064 3960 121520 0 0 3 1728 1586 139 1 13 86
0 1 0 0 3064 3964 121516 0 0 4 1856 1455 129 1 12 86
0 1 0 0 3064 3968 121512 0 0 4 2052 1717 162 0 17 83
0 1 0 0 3064 3972 121508 0 0 3 1568 1327 131 0 12 88
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 15:17 Writeout in recent kernels/VMs poor compared to last -ac Adam Kropelin
2002-01-12 15:29 ` Adam Kropelin
@ 2002-01-12 16:03 ` Rik van Riel
2002-01-12 16:30 ` Andrea Arcangeli
2002-01-12 20:47 ` Jeff Lessem
3 siblings, 0 replies; 7+ messages in thread
From: Rik van Riel @ 2002-01-12 16:03 UTC (permalink / raw)
To: Adam Kropelin; +Cc: linux-kernel
On Sat, 12 Jan 2002, Adam Kropelin wrote:
> I recently began regularly transferring large (600 MB+) files to my
> Linux-based fileserver and have noticed what I would characterize as
> poor writeout behavior under this load. I've done a bit of comparison
> testing which may help reveal the problem better.
> -ac, on the other hand, is very smooth (still a noticeable
> oscillation, but far more consistent):
Thanks for tracking down this bug. I'll try to find out what
is causing it and will attempt to fix this problem.
regards,
Rik
--
"Linux holds advantages over the single-vendor commercial OS"
-- Microsoft's "Competing with Linux" document
http://www.surriel.com/ http://distro.conectiva.com/
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 15:17 Writeout in recent kernels/VMs poor compared to last -ac Adam Kropelin
2002-01-12 15:29 ` Adam Kropelin
2002-01-12 16:03 ` Rik van Riel
@ 2002-01-12 16:30 ` Andrea Arcangeli
2002-01-12 18:19 ` Adam Kropelin
2002-01-12 20:47 ` Jeff Lessem
3 siblings, 1 reply; 7+ messages in thread
From: Andrea Arcangeli @ 2002-01-12 16:30 UTC (permalink / raw)
To: Adam Kropelin; +Cc: linux-kernel
On Sat, Jan 12, 2002 at 10:17:39AM -0500, Adam Kropelin wrote:
> I recently began regularly transferring large (600 MB+) files to my
> Linux-based fileserver and have noticed what I would characterize as poor
> writeout behavior under this load. I've done a bit of comparison testing
> which may help reveal the problem better.
>
> Disclaimer: I did not choose this test because it is scientific or stresses
> the system in any particular way. I use it because it's an operation which I
> perform frequently and so the faster it goes, the broader I smile. This may
> be an entirely inappropriate load to optimize for, and if so I understand.
>
> Test consisted of FTPing in (Linux was the destination) a 650 MB file and
> timing the transfer (from the client's perspective) which capturing "vmstat
> 1" output on the server. Server hardware is Dual PPro 200, 160 MB RAM, 512
> MB swap, destination filesystem is VFAT[0] on cpqarray partition on a RAID5
> logical drive. Root and swap are not on the cpqarray.
>
> Here are my results (average of 2 runs, each from a clean reboot with all
> unneeded services disabled):
>
> 2.4.17: 6:52
> 2.4.17-rmap11a: 6:53
> 2.4.18-pre2aa2: 7:10
> 2.4.13-ac7: 5:30 (!)
> 2.4.17 with cpqarray driver update from -ac: 6:30
>
> The last test was just to see if -ac's better performance had anything to do
> with the driver update. Apparently it had little or nothing to do with it.
>
> The vmstat output is very revealing. All tests except for -ac show a strong
> oscillation on the blocks out (this is a representative sample from stock
> 2.4.17 but the other recent kernels show essentially the same behavior):
>
> procs memory swap io system
> cpu
> r b w swpd free buff cache si so bi bo in cs us sy
> id
> 0 1 1 0 4408 2712 118756 0 0 0 0 109 16 0 0
> 100
> 0 1 1 0 4436 2716 118724 0 0 1 3913 287 13 0 8
> 92
> 0 1 0 0 4436 2716 118724 0 0 0 0 118 11 0 0
> 100
> 0 0 0 0 4352 2724 118816 0 0 8 0 3639 203 2 28
> 70
> 0 0 0 0 4420 2736 118752 0 0 11 0 4530 259 0 33
> 67
> 0 0 0 0 4348 2744 118816 0 0 10 0 4273 245 0 42
> 58
> 1 0 0 0 4376 2756 118772 0 0 11 0 4551 246 1 39
> 60
> 0 1 1 0 4364 2760 118724 0 0 4 6730 1710 93 1 19
> 80
> 0 1 1 0 4364 2760 118724 0 0 0 0 108 9 0 0
> 100
> 0 1 1 0 4364 2760 118724 0 0 0 3667 117 9 0 2
> 98
> 0 1 1 0 4364 2760 118724 0 0 0 0 125 8 0 0
> 100
> 1 0 1 0 4364 2760 118724 0 0 0 3819 124 10 0 2
> 98
> 0 1 1 0 4372 2760 118704 0 0 0 2055 195 10 0 5
> 95
> 0 1 1 0 4372 2760 118704 0 0 0 0 120 6 0 0
> 100
> 0 1 1 0 4408 2760 118668 0 0 0 2415 203 16 0 4
> 96
> 0 1 1 0 4472 2760 118608 0 0 0 4321 280 18 1 6
> 93
> 0 2 1 0 4468 2760 118608 0 0 0 0 112 8 0 0
> 100
> 0 1 1 0 4352 2760 118724 0 0 1 3232 227 9 0 10
> 90
> 0 1 1 0 4352 2760 118724 0 0 0 0 108 8 0 0
> 100
> 0 0 0 0 4440 2768 118696 0 0 7 0 3233 179 1 25
> 74
> 1 0 0 0 4448 2776 118680 0 0 11 0 4417 253 0 36
I think you simply want to trigger the soft-bdflush event earlier, with
-aa something like this may do the trick:
echo 5 500 64 256 500 3000 60 2 0 > /proc/sys/vm/bdflush
this way you'll wakeup as soon as 5% of the 118mbytes (+ free memory,
none in this case) are dirty, and bdflush will stop as soon as the level
is back to 2% (then kupdate will take care of the 2%). Those suggested
values may be too strict but this way you should get the idea if it
helps somehow or not :)
Andrea
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 16:30 ` Andrea Arcangeli
@ 2002-01-12 18:19 ` Adam Kropelin
2002-01-12 18:49 ` Jussi Laako
0 siblings, 1 reply; 7+ messages in thread
From: Adam Kropelin @ 2002-01-12 18:19 UTC (permalink / raw)
To: linux-kernel
----- Original Message -----
From: "Andrea Arcangeli" <andrea@suse.de>
To: "Adam Kropelin" <akropel1@rochester.rr.com>
Cc: <linux-kernel@vger.kernel.org>
Sent: Saturday, January 12, 2002 11:30 AM
Subject: Re: Writeout in recent kernels/VMs poor compared to last -ac
> On Sat, Jan 12, 2002 at 10:17:39AM -0500, Adam Kropelin wrote:
> > I recently began regularly transferring large (600 MB+) files to my
> > Linux-based fileserver and have noticed what I would characterize as poor
> > writeout behavior under this load. I've done a bit of comparison testing
> > which may help reveal the problem better.
<snip>
> I think you simply want to trigger the soft-bdflush event earlier, with
> -aa something like this may do the trick:
>
> echo 5 500 64 256 500 3000 60 2 0 > /proc/sys/vm/bdflush
>
> this way you'll wakeup as soon as 5% of the 118mbytes (+ free memory,
> none in this case) are dirty, and bdflush will stop as soon as the level
> is back to 2% (then kupdate will take care of the 2%). Those suggested
> values may be too strict but this way you should get the idea if it
> helps somehow or not :)
Thanks for the idea. Unfortunately, it didn't help. :(
Blocks definitely do begin hitting the disk much sooner after I begin the
transfer, but the overall time is basically unchanged: 7:08. vmstat still shows
the widely oscillating bo value.
--Adam
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 18:19 ` Adam Kropelin
@ 2002-01-12 18:49 ` Jussi Laako
0 siblings, 0 replies; 7+ messages in thread
From: Jussi Laako @ 2002-01-12 18:49 UTC (permalink / raw)
To: Adam Kropelin; +Cc: linux-kernel
Adam Kropelin wrote:
>
> Thanks for the idea. Unfortunately, it didn't help. :(
>
> Blocks definitely do begin hitting the disk much sooner after I begin the
> transfer, but the overall time is basically unchanged: 7:08. vmstat still
> shows the widely oscillating bo value.
Here's what I get with 2.4.12-ac3-lowlatency (non-fixed X11 driver):
http://www.pp.song.fi/~visitor/latencytest/3x256.html
And here's 2.4.16-lowlatency (non-fixed X11 driver):
http://www.pp.song.fi/~visitor/latencytest2/3x256.html
Here's 2.4.16-lowlatency with some small patches (fixed X11 driver):
http://www.pp.song.fi/~visitor/latencytest3/3x256.html
Here's 2.4.17-ide-rmap-O1-minill (fixed X11 driver):
http://www.pp.song.fi/~visitor/latencytest4/3x256.html
So, only usable kernel is still 2.4.12-ac3 with fixed X11 driver.
- Jussi Laako
--
PGP key fingerprint: 161D 6FED 6A92 39E2 EB5B 39DD A4DE 63EB C216 1E4B
Available at PGP keyservers
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Writeout in recent kernels/VMs poor compared to last -ac
2002-01-12 15:17 Writeout in recent kernels/VMs poor compared to last -ac Adam Kropelin
` (2 preceding siblings ...)
2002-01-12 16:30 ` Andrea Arcangeli
@ 2002-01-12 20:47 ` Jeff Lessem
3 siblings, 0 replies; 7+ messages in thread
From: Jeff Lessem @ 2002-01-12 20:47 UTC (permalink / raw)
To: linux-kernel
In your message of: Sat, 12 Jan 2002 10:17:39 EST, you write:
>I recently began regularly transferring large (600 MB+) files to my
>Linux-based fileserver and have noticed what I would characterize as poor
>writeout behavior under this load. I've done a bit of comparison testing
>which may help reveal the problem better.
I see the same behavior when writing to a USB attached harddisk/MP3
player, which also is formated VFAT. Doing a cp *.mp3 /usbdrive/ will
read the mp3s into memory, until all free memory is filled, and then
start writing them to the usb drive. Because it can only write at
800KB/s this behavior is very noticable. Once the few hundred MB of
files in memory have been written out, the next batch of files will be
read into memory, and then written out.
This behavior occurs under 2.4.17, and has occured under all recent
2.4 kernels. I haven't done any rigorous testing to see which, if
any, -ac or early 2.4 kernels do not show this behavior.
If VFAT is considered an issue, I can reformat the disk ext2 for
testing.
Not really any new information to add, just a "me too", though I would
be happy to try out any proposed fixes.
--
Jeff Lessem.
^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2002-01-12 20:48 UTC | newest]
Thread overview: 7+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2002-01-12 15:17 Writeout in recent kernels/VMs poor compared to last -ac Adam Kropelin
2002-01-12 15:29 ` Adam Kropelin
2002-01-12 16:03 ` Rik van Riel
2002-01-12 16:30 ` Andrea Arcangeli
2002-01-12 18:19 ` Adam Kropelin
2002-01-12 18:49 ` Jussi Laako
2002-01-12 20:47 ` Jeff Lessem
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.