Skip to contentSkip to conversation

Adaptive Readahead V14 - statistics question...

PATCH In review Voluspa 8 participants 29 messages May 30, 2006 – Jun 8, 2006 1 message not in the archive

Subsystems (12)
lore
Share & tools
lore

Voluspa

author
#0 lore raw

Sorry about the top-post, I'm not subscribed.

Quoted 1 line · On 2006-05-30 0:37:57 Wu Fengguang wrote:

On Mon, May 29, 2006 at 03:44:59PM -0400, Valdis Kletnieks wrote:

[...]

Quoted 9 lines
Quoted 4 lines

doing anything useful? (One thing I've noticed is that xmms, rather than gobble up 100K of data off disk every 10 seconds or so, snarfs a big 2M chunk every 3-4 minutes, often sucking in an entire song at (nearly) one shot...)

Hehe, it's resulted from the enlarged default max readahead size(128K => 1M). Too much aggressive? I'm interesting to know the recommended size for desktops, thanks. For now you can adjust it through the 'blockdev --setra /dev/hda' command.

And notebooks? I'm running a 64bit system with 2gig memory and a 7200 RPM disk. Without your patches a movie like Elephants_Dream_HD.avi causes a continuous silent read. After patching 2.6.17-rc5 (more on that later) there's a slow 'click-read-click-read-click-etc' during the same movie as the head travels _somewhere_ to rest(?) between reads.

Distracting in silent sequences, and perhaps increased disk wear/tear. I'll try adjusting the readahead size towards silence tomorrow.

But as size slides in a mainstream direction, whence will any benefit come - in this Joe-average case? It's not a faster 'cp' at least:

_Cold boot between tests - Copy between different partitions_

2.6.17-rc5-proper (Elephants_Dream_HD.avi 854537054 bytes)

real    0m44.050s
user    0m0.076s
sys     0m6.344s

2.6.17-rc5-patched

real    0m49.353s
user    0m0.075s
sys     0m6.287s

2.6.17-rc5-proper (compiled kernel tree linux-2.6.17-rc5 ~339M)

real    0m47.952s
user    0m0.198s
sys     0m6.118s

2.6.17-rc5-patched

real    0m46.513s
user    0m0.200s
sys     0m5.827s

Of course, my failure to see speed-ups could well be 'cos of a botched patch transfer (or some kind of missing groundwork only available in -mm). There was one reject in particular which made me pause. I'm no programmer... and 'continue;' is a weird direction. At the end I settled on:

[mm/readahead.c]
@@ -184,8 +289,10 @@
page->index, GFP_KERNEL)) {
ret = mapping->a_ops->readpage(filp, page); if (ret != AOP_TRUNCATED_PAGE) {
- if (!pagevec_add(&lru_pvec, page))
+ if (!pagevec_add(&lru_pvec, page)) {
+ cond_resched();
__pagevec_lru_add(&lru_pvec);
+ }
continue;
} /* else fall through to release */
}

The full 82K experiment can temprarily be found at this location:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-part-01to28of32.patch

At least it hasn't eaten my (backed up) disk yet ;-)

Mvh
Mats Johannesson
--

#1 lore raw
Quoted 16 lines · Voluspa

Sorry about the top-post, I'm not subscribed.

Quoted 1 line · On 2006-05-30 0:37:57 Wu Fengguang wrote:

On Mon, May 29, 2006 at 03:44:59PM -0400, Valdis Kletnieks wrote:

[...]

Quoted 9 lines
Quoted 4 lines

doing anything useful? (One thing I've noticed is that xmms, rather than gobble up 100K of data off disk every 10 seconds or so, snarfs a big 2M chunk every 3-4 minutes, often sucking in an entire song at (nearly) one shot...)

Hehe, it's resulted from the enlarged default max readahead size(128K => 1M). Too much aggressive? I'm interesting to know the recommended size for desktops, thanks. For now you can adjust it through the 'blockdev --setra /dev/hda' command.

Quoted 8 lines

And notebooks? I'm running a 64bit system with 2gig memory and a 7200 RPM disk. Without your patches a movie like Elephants_Dream_HD.avi causes a continuous silent read. After patching 2.6.17-rc5 (more on that later) there's a slow 'click-read-click-read-click-etc' during the same movie as the head travels _somewhere_ to rest(?) between reads.

Distracting in silent sequences, and perhaps increased disk wear/tear. I'll try adjusting the readahead size towards silence tomorrow.

Hmm... It seems risky to increase the default readahead size. I would appreciate a feed back when you are settled with some new size, thanks.

btw, maybe you will be interested in the 'laptop mode'. It prolongs battery life by making disk activity "bursty":
http://www.xs4all.nl/~bsamwel/laptop_mode/

Quoted 4 lines

But as size slides in a mainstream direction, whence will any benefit come - in this Joe-average case? It's not a faster 'cp' at least:

_Cold boot between tests - Copy between different partitions_

I have never did 'cp' tests, because it involves writes caching problems. Which makes the result hard to interpret. However I will try to explain the two tests.

Quoted 11 lines

2.6.17-rc5-proper (Elephants_Dream_HD.avi 854537054 bytes)

real    0m44.050s
user    0m0.076s
sys     0m6.344s

2.6.17-rc5-patched

real    0m49.353s
user    0m0.075s
sys     0m6.287s

- only size matters in this trivial case.
- the increased size generally do not help single reading speed.
- but it helped reducing overhead(i.e. decreased user/sys time)
- not sure why real time increased so much.

Quoted 11 lines

2.6.17-rc5-proper (compiled kernel tree linux-2.6.17-rc5 ~339M)

real    0m47.952s
user    0m0.198s
sys     0m6.118s

2.6.17-rc5-patched

real    0m46.513s
user    0m0.200s
sys     0m5.827s

- the small files optimization in the new logic helped a little

Thanks,
Wu

#2 lore raw
In reply to Voluspa
Quoted 8 lines · Voluspa
Quoted 1 line · On 2006-05-30 0:37:57 Wu Fengguang wrote:

On Mon, May 29, 2006 at 03:44:59PM -0400, Valdis Kletnieks wrote:

[...]

Quoted 5 lines
Quoted 4 lines

doing anything useful? (One thing I've noticed is that xmms, rather than gobble up 100K of data off disk every 10 seconds or so, snarfs a big 2M chunk every 3-4 minutes, often sucking in an entire song at (nearly) one shot...)

Quoted 4 lines
Quoted 4 lines

Hehe, it's resulted from the enlarged default max readahead size(128K => 1M). Too much aggressive? I'm interesting to know the recommended size for desktops, thanks. For now you can adjust it through the 'blockdev --setra /dev/hda' command.

Actually, it doesn't seem too aggressive at all - I have 768M of memory, and the larger max readahead means that it hits the disk 1/8th as often for a bigger slurp. Since I'm on a laptop with a slow 5400rpm 60g disk, a 128K seek-and-read "costs" almost exactly the same as a 1M seek-and-read...

(If I was more memory constrained, I'd probably be hitting that --setra though ;)

The only hard numbers I have so far are a build of a 17-rc4-mm3 kernel tree under -mm3+readahead and a slightly older -mm2 - the readahead kernel got through the build about 30 seconds faster (19 mins 45 secs versus 20:17 -but that's only 1 trial each).

Oh.. another "hard number" - elapsed time for a 4AM 'tripwire' run from cron with a -mm3+readahead kernel was 36 minutes. A few days earlier, a -mm3 kernel took 46 minutes for the same thing. I'll have to go and retry this with equivalent cache-cold scenarios - I *think* the file cache was roughly equivalent, but can't prove it...

The desktop "feel" is certainly at least as good, but it's a lot harder to quantify that - yesterday I was doing some heavy-duty cleaning in my ~/Mail directory (MH-style one message per file, about 250K files and 3G, obviously seriously in need of cleaning). I'd often have 2 different 'find | xargs grep' type commands running at a time, and that seemed to work a lot better than it used to (but again, no numbers)..

Damn, this is a lot harder to benchmark than the sort of microbenchmarks we usually see around here. :)

Quoted 5 lines

And notebooks? I'm running a 64bit system with 2gig memory and a 7200 RPM disk. Without your patches a movie like Elephants_Dream_HD.avi causes a continuous silent read. After patching 2.6.17-rc5 (more on that later) there's a slow 'click-read-click-read-click-etc' during the same movie as the head travels _somewhere_ to rest(?) between reads.

For my usage patterns, this is a feature, not a bug. As mentioned before, on this machine anything that reduces the number of seeks is a Good Thing.

Quoted 1 line

Distracting in silent sequences, and perhaps increased disk wear/tear.

It would be increased wear/tear only if the disk was idle long enough to spin down. Especially for video, the read-ahead needed to let the disk spin down (assuming a sane timeout for that) would be enormous. :)

Quoted 1 line

I'll try adjusting the readahead size towards silence tomorrow.

The onboard sound chip is an ok-quality CS4205, the onboard speakers are crap. However, running the audio through a nice pair of Kenwood headphones is a good solution. I don't hear the disk (or sometimes even the phone), and my co-workers don't have to hear my Malmsteen collection. :)

#3 lore raw
Quoted 2 lines · El Tue, 30 May 2006 12:49:50 -0400, [email protected] escribió:

The desktop "feel" is certainly at least as good, but it's a lot harder to quantify that - yesterday I was doing some heavy-duty cleaning in my

My desktop seems to boot a bit faster with adaptive readahead. I setup a environment running kdm with automatic login plus a kde session which runs a konqueror window and a openoffice writer windows. The time it takes for the system to show the OO window went from 1:19 to 1:16 (I did a couple of test of each kernel). Not a very scientific measurement, bootchart probably could do it better

author
#4 lore raw
In reply to [email protected]
Quoted 3 lines · On Tue, 30 May 2006 12:49:50 -0400 Valdis.Kletnieks wrote:
Quoted 2 lines · Voluspa
Quoted 1 line · On 2006-05-30 0:37:57 Wu Fengguang wrote:

On Mon, May 29, 2006 at 03:44:59PM -0400, Valdis Kletnieks wrote:

[...]

Quoted 2 lines

Damn, this is a lot harder to benchmark than the sort of microbenchmarks we usually see around here. :)

I don't even know what a microbenchmark is, but 'cp' and its higher-level equivalents is such a frequent operation that I always begin any test there.

[...] [Correction, should be: 'click-read-pause, click-read-pause etc']

Quoted 11 lines
Quoted 2 lines

later) there's a slow 'click-read-click-read-click-etc' during the same movie as the head travels _somewhere_ to rest(?) between reads.

For my usage patterns, this is a feature, not a bug. As mentioned before, on this machine anything that reduces the number of seeks is a Good Thing.

Quoted 1 line

Distracting in silent sequences, and perhaps increased disk wear/tear.

It would be increased wear/tear only if the disk was idle long enough to spin down. Especially for video, the read-ahead needed to let the disk spin down (assuming a sane timeout for that) would be enormous. :)

:-) I was thinking more in terms of disk head _arm_ wear. Somehow there's a picture in my head of the arm swinging back to a rest position at an outer (or inner?) "safe" disk track if read/write operations are delayed too much. And therefore I associate a 'click' with the arm swinging back into action. Normal quick read/write arm movement noise is distinctly different - in my uninformed user ears.

I haven't adjusted the readahed size yet, but instead performed a series of real-world usage tests.

Conclusion: On _this_ machine, with _these_ operations, Adaptive Readahead in its current incarnation and default settings is a _loss_.

Patch version:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-part-01to28of32-and-update-01to04of04-and-saner-CDVD-medium-error-handling.patch

Relevant hardware:
AMD Athlon 64 Processor 3400+ (2200 MHz top speed) L1 I Cache: 64K (64 bytes/line), D cache 64K (64 bytes/line), L2 Cache: 1024K (64 bytes/line). VIA K8M800 chipset with VT8235 south. Kingmax 2x1GB DDR-333MHz SO-DIMM memory. Hitachi Travelstar 7K100 (HTS721010G9AT00) 100GB 7200RPM Parallel-ATA disk,
http://www.hitachigst.com/hdd/support/7k100/7k100.htm acoustic management value was set to 254 (fast/"noisy") at delivery.

Soft system:
Is extremely lean and simple. Pure 64bit compiled in a lfs-ish way almost exactly 1 year ago. No desktop, just a wm (which wasn't even launched in these tests). Toolchain glibc-2.3.5 (nptl), binutils-2.16.1, gcc-3.4.4

Filesystem:
Journaled ext3 with default mount (ordered data mode) and noatime.

Kernels:

loke:sleipner:~$ ls -l /boot/kernel-2.6.17-rc5*
1440 -rw-r--r--  1 root root 1469211 May 30 02:25 /boot/kernel-2.6.17-rc5
1444 -rw-r--r--  1 root root 1470540 May 30 19:07 /boot/kernel-2.6.17-rc5-ar

All tests were performed as the root user from a machine standstill "cold boot" for each iteration and prepared for a 'console login - immediate run' ie. eventual previous output deleted/reset.

_Massive READ_

[/usr had some 490000 files]

"cd /usr ; time find . -type f -exec md5sum {} \;"

2.6.17-rc5 ------- 2.6.17-rc5-ar

real 21m21.009s -- 21m37.663s
user 3m20.784s -- 3m20.701s
sys 6m34.261s -- 6m41.735s

I had planned to run this at least three times, but didn't realize I had 12 compiled kernel trees and 3 uncompiled there... So, a one-shot had to do. But it's still significant.

_READ/WRITE_

[255 .tga files, each is 1244178 bytes]
[1 .wav file which is 1587644 bytes]
[movie becomes 573298 bytes ~9s long]

"time mencoder -ovc lavc -lavcopts aspect=16/9 mf://picsave/kreation/03-logo-joined/*.tga -oac lavc -audiofile kreation-files/kreation-logo-final.wav -o logo-final-widescreen-speedtest.avi"

2.6.17-rc5

real 0m10.164s 0m10.224s 0m10.141s

user 0m3.301s  0m3.304s  0m3.297s
sys  0m1.103s  0m1.097s  0m1.082s

2.6.17-rc5-ar

real 0m10.831s 0m10.816s 0m10.747s

user 0m3.319s  0m3.313s  0m3.324s
sys  0m1.081s  0m1.099s  0m1.042s

A 0.6s slowdown might not seem as such a big deal, but this is on a 9s movie! Furthermore, the test was conducted on the / root partition which resides on hda2. Subtracting the 8GB hda1 and the occupied 1.2GB of hda2 places us 9.2GB in from the disk edge (assuming 1 platter). I did a one-shot test of this movie on hda3 - closes to the spindle - which all in all gives a distance of ~95GB:

2.6.17-rc5 ------ 2.6.17-rc5-ar

real 0m16.134s -- 0m17.456s
user 0m3.311s -- 0m3.312s
sys 0m1.111s -- 0m1.135s

Wow. If nothing else, these tests have made me rethink my partitioning scheme. I've used the same layout since xx-years ago when proximity of swap-usr-home on those slow disks really made a difference. And since I don't touch swap in normal operation nowadays... Power to the Edge!

_Geek usage_

[Kernel compile]
[CONFIG_REORDER "Processor type and features -> Function reordering" adds ca 30s here]
[Note: I made a mistake by booting the -ar kernel first, and also didn't alternate like I should have. This was the first set of tests and chip temperature rise seem to slow things down. Physics reason above my head]

"time make"

2.6.17-rc5-ar

real 5m3.654s 5m3.787s 5m4.390s 5m4.991s
user 4m17.595s 4m17.580s 4m17.701s 4m18.043s
sys 0m31.551s 0m31.506s 0m31.368s 0m31.563s

2.6.17-rc5

real 5m4.606s 5m5.798s 5m4.684s 5m4.508s
user 4m18.586s 4m19.183s 4m19.111s 4m17.799s
sys 0m31.241s 0m31.482s 0m31.278s 0m31.610s

Any difference here should really be considered noise. The file read/write is too infrequent and slow to really measure.

_Caveat and preemptive Mea Culpa_

The patching of 2.6.17-rc5 has neither been approved or verified as to its correctness. The kernel compiles without errors and the new /proc/sys/kernel/ sysctl readahead_ratio and readahead_hit_rate turn up with the default 50 and 1. This is however not a proof of total parity with the official -mm patch-set.

Mvh
Mats Johannesson
--

#5 lore raw
Quoted 8 lines · Voluspa

_Massive READ_

[/usr had some 490000 files]

"cd /usr ; time find . -type f -exec md5sum {} \;"

2.6.17-rc5 ------- 2.6.17-rc5-ar

Quoted 7 lines

real 21m21.009s -- 21m37.663s
user 3m20.784s -- 3m20.701s
sys 6m34.261s -- 6m41.735s

I had planned to run this at least three times, but didn't realize I had 12 compiled kernel trees and 3 uncompiled there... So, a one-shot had to do. But it's still significant.

Sorry, it is a known regression. I'd like to fix it in the next release.

Thanks,
Wu

author
#6 lore raw
Quoted 13 lines · On Thu, 1 Jun 2006 13:51:43 +0800 Fengguang Wu wrote:
Quoted 12 lines · Voluspa

_Massive READ_

[/usr had some 490000 files]

"cd /usr ; time find . -type f -exec md5sum {} \;"

2.6.17-rc5 ------- 2.6.17-rc5-ar

real 21m21.009s -- 21m37.663s
user 3m20.784s -- 3m20.701s
sys 6m34.261s -- 6m41.735s

Quoted 6 lines
Quoted 3 lines

I had planned to run this at least three times, but didn't realize I had 12 compiled kernel trees and 3 uncompiled there... So, a one-shot had to do. But it's still significant.

Sorry, it is a known regression. I'd like to fix it in the next release.

That's cool. I had fun testing (I'm weird) and now have a fixed procedure to monitor your future work. When/if it hits mainline I'll both back it out and switch it on/off. Then shout WOLF if I see a regression anywhere ;)

There's still the readahead size to adjust. I'll return with my findings.

Mvh
Mats Johannesson
--

author
#7 lore raw
In reply to Fengguang Wu

My patching was borked as can be seen in:
http://marc.theaimsgroup.com/?l=linux-kernel&m=114956084026066&w=2

I've therefore benchmarked two corrected patches from Wu:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-2.6.17-rc5-wu-v1.patch
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-2.6.17-rc5-wu-v2.patch

Revised Conclusion: On _this_ machine, with _these_ operations, Adaptive Readahead in its current incarnation and default settings is a slight _loss_. However, if the readahead size is lowered from 2048 to 256, it becomes a slight _gain_ or at least stays in parity with normal readahead.

I suggest others to test multi-thread, multi-cpu, more than/less than my 2GB memory, sata-disks, different disk speeds etc etc.

Kernels:

root:sleipner:~# ls -l /boot/kernel-2.6.17-rc5-git10*
1440 -rw-r--r--  1 root root 1469326 Jun  6 22:27 /boot/kernel-2.6.17-rc5-git10
1440 -rw-r--r--  1 root root 1470122 Jun  6 22:36 /boot/kernel-2.6.17-rc5-git10-ar1
1440 -rw-r--r--  1 root root 1470128 Jun  6 22:44 /boot/kernel-2.6.17-rc5-git10-ar2

_Massive READ_

[/usr had some 195000 files]

"cd /usr; time find . -type f -exec md5sum {} \; >/dev/null"

[/sbin/blockdev --setra 256 /dev/hda] * [/sbin/blockdev --setra 2048 /dev/hda]

6.17-rc5-git10 - git10-ar1 - git10-ar2 * 6.17-rc5-git10 - git10-ar1 - git10-ar2

real 8m18.241s - 8m19.053s - 8m16.639s * real 8m24.042s - 8m22.652s - 8m20.812s user 1m23.556s - 1m24.526s - 1m23.725s * user 1m23.788s - 1m23.741s - 1m24.023s sys 2m8.514s - 2m5.989s - 2m3.540s * sys 2m7.369s - 2m6.914s - 2m5.317s

real 8m19.171s - 8m17.993s - 8m17.062s * real 8m23.110s - 8m20.409s - 8m19.278s user 1m23.863s - 1m23.692s - 1m23.980s * user 1m23.770s - 1m23.715s - 1m23.525s sys 2m9.332s - 2m4.133s - 2m3.602s * sys 2m6.463s - 2m5.735s - 2m3.801s

real 8m17.111s - 8m17.102s - 8m16.859s * real 8m21.891s - 8m19.129s - 8m17.321s user 1m24.071s - 1m24.126s - 1m24.430s * user 1m23.876s - 1m23.592s - 1m23.024s sys 2m6.292s - 2m3.543s - 2m3.142s * sys 2m4.768s - 2m4.012s - 2m3.110s

real 8m20.427s - 8m16.972s - 8m17.847s * real 8m25.359s - 8m21.261s - 8m20.365s user 1m23.730s - 1m23.260s - 1m23.227s * user 1m24.242s - 1m23.825s - 1m23.895s sys 2m9.524s - 2m3.708s - 2m5.244s * sys 2m7.894s - 2m5.366s - 2m3.971s

_READ/WRITE_

[255 .tga files, each is 1244178 bytes]
[1 .wav file which is 1587644 bytes]
[movie becomes 573298 bytes ~9s long]

"time mencoder -ovc lavc -lavcopts aspect=16/9 mf://picsave/kreation/03-logo-joined/*.tga -oac lavc -audiofile kreation-files/kreation-logo-final.wav -o logo-final-widescreen-speedtest.avi"

[/sbin/blockdev --setra 256 /dev/hda] * [/sbin/blockdev --setra 2048 /dev/hda]

6.17-rc5-git10 - git10-ar1 - git10-ar2 * 6.17-rc5-git10 - git10-ar1 - git10-ar2

real 0m12.961s - 0m12.864s - 0m12.811s * real 0m16.628s - 0m15.862s - 0m16.754s

user 0m3.315s  - 0m3.319s  - 0m3.316s  * user 0m3.335s  - 0m3.301s  - 0m3.308s
sys  0m1.082s  - 0m1.077s  - 0m1.086s  * sys  0m1.084s  - 0m1.122s  - 0m1.093s

real 0m12.908s - 0m12.793s - 0m12.813s * real 0m16.601s - 0m15.893s - 0m16.736s

user 0m3.323s  - 0m3.305s  - 0m3.312s  * user 0m3.311s  - 0m3.316s  - 0m3.308s
sys  0m1.051s  - 0m1.079s  - 0m1.145s  * sys  0m1.046s  - 0m1.109s  - 0m1.091s

_cp bigfile between different partitions_

[Elephants_Dream_HD.avi 854537054 bytes]

"time cp /home/downloads/Elephants_Dream_HD.avi /root"

[/sbin/blockdev --setra 256 /dev/hda] * [/sbin/blockdev --setra 2048 /dev/hda]

6.17-rc5-git10 - git10-ar1 - git10-ar2 * 6.17-rc5-git10 - git10-ar1 - git10-ar2

real 0m46.463s - 0m46.909s - 0m45.865s * real 0m50.232s - 0m50.863s - 0m50.549s

user 0m0.081s  - 0m0.073s  - 0m0.068s  * user 0m0.069s  - 0m0.063s  - 0m0.088s
sys  0m6.304s  - 0m7.204s  - 0m5.949s  * sys  0m5.902s  - 0m7.174s  - 0m6.822s

real 0m46.126s - 0m47.305s - 0m47.174s * real 0m50.875s - 0m50.066s - 0m50.862s

user 0m0.091s  - 0m0.095s  - 0m0.070s  * user 0m0.099s  - 0m0.091s  - 0m0.071s
sys  0m5.751s  - 0m7.159s  - 0m6.707s  * sys  0m6.271s  - 0m6.740s  - 0m7.318s

_cp filetree between different partitions_

[compiled kerneltree ~339M]

"time cp -a /usr/src/testing/linux-2.6.17-rc5-git10 /root"

[/sbin/blockdev --setra 256 /dev/hda] * [/sbin/blockdev --setra 2048 /dev/hda]

6.17-rc5-git10 - git10-ar1 - git10-ar2 * 6.17-rc5-git10 - git10-ar1 - git10-ar2

real 0m51.344s - 0m51.886s - 0m51.077s * real 0m52.502s - 0m52.757s - 0m54.794s

user 0m0.193s  - 0m0.220s  - 0m0.231s  * user 0m0.210s  - 0m0.198s  - 0m0.177s
sys  0m5.508s  - 0m6.003s  - 0m5.205s  * sys  0m5.980s  - 0m5.800s  - 0m6.372s

real 0m51.148s - 0m51.212s - 0m51.768s * real 0m51.488s - 0m52.098s - 0m51.719s

user 0m0.170s  - 0m0.209s  - 0m0.184s  * user 0m0.198s  - 0m0.210s  - 0m0.179s
sys  0m5.697s  - 0m5.604s  - 0m6.438s  * sys  0m5.527s  - 0m5.918s  - 0m5.544s

Mvh
Mats Johannesson
--

#8 lore raw

adaptive readahead overheads

I'd like to show some numbers on the pure software overheads come with the adaptive readahead in daily operations.

SEQUENTIAL ACCESS OVERHEADS, or: PER-PAGE OVERHEADS
===================================================

SUMMARY
user sys cpu total
ARA 0.13 5.30 92.0% 5.87
STOCK 0.12 5.34 91.8% 5.91

It shows about 0.8% overhead.

They are mainly contributed by:
- debug/accounting code
- smooth aging accounting for the stateful method
- readahead hit feedback

However, if necessary, each of the obove can be seperated out and made a kconfig option. They won't affect the main functionality of ARA.

DETAILS

% time cp work/sparse /dev/null # repeat 5 times each

/proc/sys/vm/readahead_ratio = 1

cp work/sparse /dev/null  0.13s user 5.30s system 93% cpu 5.780 total
cp work/sparse /dev/null  0.13s user 5.29s system 92% cpu 5.847 total
cp work/sparse /dev/null  0.13s user 5.27s system 92% cpu 5.860 total
cp work/sparse /dev/null  0.13s user 5.32s system 91% cpu 5.961 total
cp work/sparse /dev/null  0.12s user 5.32s system 92% cpu 5.902 total

/proc/sys/vm/readahead_ratio = 100

cp work/sparse /dev/null  0.12s user 5.39s system 93% cpu 5.888 total
cp work/sparse /dev/null  0.13s user 5.32s system 92% cpu 5.923 total
cp work/sparse /dev/null  0.11s user 5.32s system 91% cpu 5.901 total
cp work/sparse /dev/null  0.12s user 5.35s system 92% cpu 5.952 total
cp work/sparse /dev/null  0.12s user 5.30s system 91% cpu 5.900 total

SMALL FILES OVERHEADS, or: PER-FILE OVERHEADS
=============================================

SUMMARY
user sys cpu total
ARA 405.90 326.54 97% 12:27.59
stock 407.53 322.02 97% 12:27.52

No obvious overhead.

There is overhead in calling readahead_close() on each file. However, the small-io-go-all-down-to-lowlevel events are sucessfully reduced, making up for the overhead.

DETAILS
In a qemu with 156M memory and a host system with 4G memory, traverse a whole tree of 434M /usr, and do this repeatedly. So that the 2+ runs do not involve true disk I/O.

# time find /usr -type f -exec md5sum {} \; >/dev/null

ARA

406.00s user 325.16s system 97% cpu 12:28.17 total 403.35s user 325.15s system 97% cpu 12:23.86 total 403.00s user 325.03s system 97% cpu 12:23.61 total 406.43s user 327.64s system 97% cpu 12:30.64 total 407.03s user 325.46s system 97% cpu 12:28.17 total 406.46s user 326.89s system 98% cpu 12:26.31 total 405.96s user 328.45s system 98% cpu 12:27.08 total 409.05s user 328.55s system 97% cpu 12:32.85 total

STOCK(vanilla kernel)

408.64s user 321.55s system 97% cpu 12:27.68 total 406.39s user 320.55s system 97% cpu 12:24.58 total 408.41s user 321.22s system 97% cpu 12:27.91 total 409.10s user 324.72s system 97% cpu 12:33.16 total 406.60s user 321.81s system 97% cpu 12:26.48 total 405.60s user 322.49s system 97% cpu 12:25.29 total 408.13s user 322.19s system 97% cpu 12:27.92 total 407.35s user 321.66s system 97% cpu 12:27.11 total

Thanks,
Wu

author
#9 lore raw

Re: adaptive readahead overheads

Quoted 2 lines · On Thu, 8 Jun 2006 19:37:31 +0800 Fengguang Wu wrote:

I'd like to show some numbers on the pure software overheads come with the adaptive readahead in daily operations.

[...]

Quoted 6 lines
# time find /usr -type f -exec md5sum {} \; >/dev/null

ARA

406.00s user 325.16s system 97% cpu 12:28.17 total

Just out of interest, all your figures show an almost maxed out CPU. Why is it that my own runs use so little CPU? I'm running the above command as we 'speak' and on average only 40% is utilized, with the occasional spike at max 75%. And this is on the lowest CPU level 800MHz, which means that the 80% up_threshold of the ondemand cpufreq governor never is breached (there are 1800MHz, 2000MHz and 2200MHz levels above it).

Are you only 'giving' qemu something like 400MHz to play with or is qemu so inefficient in itself?

Mvh
Mats Johannesson
--

#10 lore raw

Re: adaptive readahead overheads

Quoted 5 lines · Voluspa
Quoted 2 lines · On Thu, 8 Jun 2006 19:37:31 +0800 Fengguang Wu wrote:

I'd like to show some numbers on the pure software overheads come with the adaptive readahead in daily operations.

[...]

Quoted 1 line
Quoted 8 lines
Quoted 5 lines
# time find /usr -type f -exec md5sum {} \; >/dev/null

ARA

406.00s user 325.16s system 97% cpu 12:28.17 total

Just out of interest, all your figures show an almost maxed out CPU. Why is it that my own runs use so little CPU? I'm running the above

It does not have to wait for disk _seeks_, I guess.
All reads inside qemu actually hit the page cache in the host system ;)

Quoted 8 lines

command as we 'speak' and on average only 40% is utilized, with the occasional spike at max 75%. And this is on the lowest CPU level 800MHz, which means that the 80% up_threshold of the ondemand cpufreq governor never is breached (there are 1800MHz, 2000MHz and 2200MHz levels above it).

Are you only 'giving' qemu something like 400MHz to play with or is qemu so inefficient in itself

My qemu command line is:

        qemu -m 156 /lab/wfg/linux_image -kernel ./arch/i386/boot/bzImage
                        -append "root=/dev /hda console=ttyS0,9600" -nographic

I do not have the qemu accelerator, though.

Thanks,
Wu

#11 lore raw

Re: adaptive readahead overheads

In reply to Voluspa
Quoted 5 lines · Voluspa
Quoted 2 lines · On Thu, 8 Jun 2006 19:37:31 +0800 Fengguang Wu wrote:

I'd like to show some numbers on the pure software overheads come with the adaptive readahead in daily operations.

[...]

Quoted 1 line
Quoted 7 lines
Quoted 5 lines
# time find /usr -type f -exec md5sum {} \; >/dev/null

ARA

406.00s user 325.16s system 97% cpu 12:28.17 total

Just out of interest, all your figures show an almost maxed out CPU.

It might be because qemu has a poor IO model (old IDE) that is quite CPU intensive to drive.

-Andi

#12 lore raw

Re: adaptive readahead overheads

Quoted 9 lines · Andi Kleen
Quoted 7 lines · Voluspa
Quoted 2 lines · On Thu, 8 Jun 2006 19:37:31 +0800 Fengguang Wu wrote:

I'd like to show some numbers on the pure software overheads come with the adaptive readahead in daily operations.

[...]

Quoted 3 lines
# time find /usr -type f -exec md5sum {} \; >/dev/null
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

ARA

406.00s user 325.16s system 97% cpu 12:28.17 total

Just out of interest, all your figures show an almost maxed out CPU.

It might be because qemu has a poor IO model (old IDE) that is quite CPU intensive to drive.

qemu ide supports dma, so it's ok in that respect. It doesn't support async io for the host though, so all io ends up blocking waiting for io to complete at the host. I would not advocate using qemu for benching these things, it's just not (yet) well suited for io testing.

Signature · 1 line
Jens Axboe

A message that is not in the archive (sent off-list or not yet received): [email protected]

#14 lore raw
Quoted 33 lines · Bart Samwel

Hi Mats, Wu,

Hmmm, video at 1 Mb/s = 128 kB/s (just guessing a typical bitrate) equals 8 seconds between reads at 1 MB readahead, right? That's strange, you should be hearing those sounds normally as well then, as a typical Linux laptop setup accesses the disk less frequently than once every 8 seconds. Anyway, _increasing_ the maximum readahead to some film-decent value will probably get rid of the clicking as well.

I guess the problem is getting this to work without disturbing other applications, i.e. without making slow-but-predictably-reading applications read ahead 10 MB as well. I've been struggling with this with laptop mode for quite some time: last time I checked, there didn't seem to be a good way to do video readahead without making all other reads read ahead too much as well... What I'd like to have is a setting that works based on _time_, so that I can say "read all you think will be needed in the next N seconds".

I could imagine having the maximum readahead being composed of two settings:

MAX_BYTES = maximum readahead in bytes
MAX_TIME = maximum readahead in *time* before the data is expected to be needed

For instance, if MAX_BYTES = 50MB and MAX_TIME=180 seconds, an application reading at 10 kB/s would get a max readahead of 180*10 = 1800kB, while an application reading at 100 kB/s would get a max readahead of 180*100 = 18000kB. As a use case, the first application would be xmms (128kbit MP3), while the second would be mplayer (800kbit video). In both cases, laptop mode would be able to spin down the disk for a full three minutes between spinups. Ideal for when you're trying to watch a video while on the road.

Quoted 2 lines

Wu, do the adaptive readahead patches have something like this, or could it be included? It would solve a _major_ problem for laptop mode.

Yes, it has the capability you need.

And MAX_TIME is not necessary for this case.
It does not try to estimate the time, but the relative speeds of all concurrent readers. It automatically arranges appropriate readahead sizes for all the concurrent readers, so that no readahead thrashing will happen, provided that the reading speeds do not fluctuate too much.

However there are some cases not thrashing protected:
- normal mmapped reading(without POSIX_FADV_SEQUENTIAL hint);
- backward reading;
- fs stuffs(i.e. readahead for dirs);

With that in mind, you can safely set max readahead size to as large as 255M when watching videos by doing the following tricks:

blockdev --setra 524280 /dev/hda[N]     # following opened file will use this aggressive size
mplayer <your video file on hda[N]>     # open it
blockdev --setra 2048 /dev/hda[N]       # revert to sane value
# now continue watching video ...

Cheers,
Wu

#15 lore raw

PATCH 1/1 readahead: initial method - expected read size - fix fastcall

Remove 'fastcall' directive for function readahead_close().

It has drawn concerns from Andrew Morton. Now I have some benchmarks on it, and proved it as a _false_ optimization.

The tests are simple runs of the following command over _cached_ dirs:
time find / > /dev/null

Table of summary(averages):

		user		sys		cpu		total

fastcall: 1.236 4.39 89% 6.2936
non-fastcall: 1.18 4.14166667 92% 5.75416667
stock: 1.25833333 4.14666667 93.3% 5.75866667

-----------
Detailed outputs:

readahead patched kernel with fastcall:

noglob find / > /dev/null  1.21s user 4.58s system 90% cpu 6.378 total
noglob find / > /dev/null  1.25s user 4.47s system 86% cpu 6.623 total
noglob find / > /dev/null  1.23s user 4.36s system 90% cpu 6.173 total
noglob find / > /dev/null  1.25s user 4.33s system 92% cpu 6.067 total
noglob find / > /dev/null  1.24s user 4.21s system 87% cpu 6.227 total

readahead patched kernel without fastcall:

noglob find / > /dev/null  1.21s user 4.46s system 95% cpu 5.962 total
noglob find / > /dev/null  1.26s user 4.58s system 94% cpu 6.142 total
noglob find / > /dev/null  1.10s user 3.80s system 86% cpu 5.661 total
noglob find / > /dev/null  1.13s user 3.98s system 95% cpu 5.355 total
noglob find / > /dev/null  1.18s user 4.00s system 89% cpu 5.805 total
noglob find / > /dev/null  1.22s user 4.03s system 93% cpu 5.600 total

stock kernel:

noglob find / > /dev/null  1.22s user 4.24s system 94% cpu 5.803 total
noglob find / > /dev/null  1.31s user 4.21s system 95% cpu 5.784 total
noglob find / > /dev/null  1.27s user 4.24s system 97% cpu 5.676 total
noglob find / > /dev/null  1.34s user 4.21s system 94% cpu 5.844 total
noglob find / > /dev/null  1.26s user 4.08s system 89% cpu 5.935 total
noglob find / > /dev/null  1.15s user 3.90s system 91% cpu 5.510 total

-----------
Similar regression has also been found by Voluspa <[email protected]>:

Quoted 7 lines

"cd /usr ; time find . -type f -exec md5sum {} \;"

2.6.17-rc5 ------- 2.6.17-rc5-ar

real 21m21.009s -- 21m37.663s
user 3m20.784s -- 3m20.701s
sys 6m34.261s -- 6m41.735s

---

include/linux/mm.h +1 -1
@@ -1068,7 +1068,7 @@ unsigned long page_cache_readahead(struc
 void handle_ra_miss(struct address_space *mapping, 
 		    struct file_ra_state *ra, pgoff_t offset);
 unsigned long max_sane_readahead(unsigned long nr);
-void fastcall readahead_close(struct file *file);
+void readahead_close(struct file *file);
 unsigned long
 page_cache_readahead_adaptive(struct address_space *mapping,
 			struct file_ra_state *ra, struct file *filp,
mm/readahead.c +1 -1
@@ -1943,7 +1943,7 @@ void fastcall readahead_cache_hit(struct
  * The resulted `ra_expect_bytes' answers the question of:
  * 	How many pages are expected to be read on start-of-file?
  */
-void fastcall readahead_close(struct file *file)
+void readahead_close(struct file *file)
 {
 	struct inode *inode = file->f_dentry->d_inode;
 	struct address_space *mapping = inode->i_mapping;
#16 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 3 lines · On Sun, 4 Jun 2006 15:34:15 +0800 Fengguang Wu <[email protected]> wrote:

Remove 'fastcall' directive for function readahead_close().

It has drawn concerns from Andrew Morton.

Well. I think fastcall is ugly and vaguely silly. Now if we has a really_really_fastcall then I'd like to use that!

Quoted 2 lines

Now I have some benchmarks
on it, and proved it as a _false_ optimization.

Sorry, I don't believe this will be measurable (and with CONFIG_REGPARM it'll be a no-op).

But I'm always glad to see a fastcall disappear ;)

#17 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 11 lines · Andrew Morton
Quoted 3 lines · On Sun, 4 Jun 2006 15:34:15 +0800 Fengguang Wu <[email protected]> wrote:

Remove 'fastcall' directive for function readahead_close().

It has drawn concerns from Andrew Morton.

Well. I think fastcall is ugly and vaguely silly. Now if we has a really_really_fastcall then I'd like to use that!

Quoted 5 lines
Quoted 2 lines

Now I have some benchmarks
on it, and proved it as a _false_ optimization.

Sorry, I don't believe this will be measurable (and with CONFIG_REGPARM it'll be a no-op).

we should just make CONFIG_REGPARM be "it" always (and thus make it go away as config option) and then just remove all "fastcall" from the kernel...

author
#18 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 15 lines · On Sun, 04 Jun 2006 11:25:03 +0200 Arjan van de Ven wrote:
Quoted 14 lines · Andrew Morton
Quoted 3 lines · On Sun, 4 Jun 2006 15:34:15 +0800 Fengguang Wu wrote:

Remove 'fastcall' directive for function readahead_close().

It has drawn concerns from Andrew Morton.

Well. I think fastcall is ugly and vaguely silly. Now if we has a really_really_fastcall then I'd like to use that!

Quoted 2 lines

Now I have some benchmarks
on it, and proved it as a _false_ optimization.

Quoted 6 lines
Quoted 2 lines

Sorry, I don't believe this will be measurable (and with CONFIG_REGPARM it'll be a no-op).

we should just make CONFIG_REGPARM be "it" always (and thus make it go away as config option) and then just remove all "fastcall" from the kernel...

Wu, I don't know anything about REGPARM, which my x86_64 config doesn't have, (only the never kernel-updated i686 machine) but your last 2 patches - I didn't apply the, to me, seemingly unrelated radixtree thing - sure is measurable. In a most negative way regarding READ/WRITE and a completely unchanged READ.

Patch:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-git-updated-june-04-2006.patch

Kernel config:
http://web.comhem.se/~u46139355/storetmp/voluspa-2.6.17-rc5-git10-ar-.config

Hard/soft-ware and testing procedure:
http://marc.theaimsgroup.com/?l=linux-kernel&m=114911241320301&w=2

I should add that CFQ is the IO scheduler used and what I mean by default adaptive readahead options is just that, minus the debug stuff (not compiled). Also, I've verified that the CFQ updates in -git10 didn't regress plain -rc5.

_Massive READ_

[/usr had some 166000 files]

By routing screen writes to /dev/null, this test became so stable that a one-shot would have sufficed, but I did two for those who crave pudding.

"cd /usr; time find . -type f -exec md5sum {} \; >/dev/null"

2.6.17-rc5-git10 ---------- 2.6.17-rc5-git10-ar

real 7m15.525s 7m15.425s -- 7m19.720s 7m19.065s
user 1m11.596s 1m11.894s -- 1m12.111s 1m11.585s
sys 1m49.211s 1m48.068s -- 1m46.223s 1m46.887s

When the target had 490000 files (and wrote to screen) the difference was 16.5 seconds, so the 4s here is in line (time hovered at 7m34s for -git10 and 7m39s for -git10-ar when writing to screen, +/- 1-2s).

_READ/WRITE_

[255 .tga files, each is 1244178 bytes]
[1 .wav file which is 1587644 bytes]
[movie becomes 573298 bytes ~9s long]

This is also extremely accurate _between runs_. Time varies depending on file order and placement on disk, but the actual runs are 'klockrena' ('dead accurate'?)

time mencoder -ovc lavc -lavcopts aspect=16/9 mf://picsave/kreation/03-logo-joined/*.tga -oac lavc -audiofile kreation-files/kreation-logo-final.wav -o logo-final-widescreen-speedtest.avi

2.6.17-rc5-git10 ---------- 2.6.17-rc5-git10-ar

real 0m12.892s 0m12.903s -- 0m16.555s 0m16.611s

user 0m3.289s  0m3.314s  -- 0m3.315s  0m3.307s
sys  0m1.124s  0m1.085s  -- 0m1.096s  0m1.121s

Without the 2 new patches, the difference was 0.6s, still bad on a 9s movie, but these 3.5s are flabbergasting!

And, no. I don't know which of the 2 that caused this. I am prepared to hand out a link to a .bz2 archive of all the files involved. Just send a private message. You only need mplayer [mencoder] installed, and then a one-shot tells it all.

Mvh
Mats Johannesson
--

#19 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 16 lines · Voluspa
Quoted 15 lines · On Sun, 04 Jun 2006 11:25:03 +0200 Arjan van de Ven wrote:
Quoted 14 lines · Andrew Morton
Quoted 3 lines · On Sun, 4 Jun 2006 15:34:15 +0800 Fengguang Wu wrote:

Remove 'fastcall' directive for function readahead_close().

It has drawn concerns from Andrew Morton.

Well. I think fastcall is ugly and vaguely silly. Now if we has a really_really_fastcall then I'd like to use that!

Quoted 2 lines

Now I have some benchmarks
on it, and proved it as a _false_ optimization.

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Sorry, I don't believe this will be measurable (and with CONFIG_REGPARM it'll be a no-op).

we should just make CONFIG_REGPARM be "it" always (and thus make it go away as config option) and then just remove all "fastcall" from the kernel...

Wu, I don't know anything about REGPARM, which my x86_64 config doesn't have,

because it doesn't need it since it's default for that architecture

#20 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

In reply to Voluspa
Quoted 2 lines · Voluspa

Patch:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-git-updated-june-04-2006.patch

It seems that the patch has some problem:

+ /* ra_size in its _steady_ state reflects thrashing threshold */
+ if (page && ra_old + ra_old / 8 >= ra_size)
+ update_ra_thrash_bytes(mapping->backing_dev_info, ra_size);
+
+ limit_rala(growth_limit, la_old, &ra_size, &la_size);

The above statements was displaced, rendering the if() clause to fail all the time. That defeats the small file optimization, for ra_thrash_bytes will remain small.

Voluspa, I attached an updated patch, including two more performance tunings. Please take it when you find time to do more benchmarks, thanks.

I'd suggest that you(and other kind people interested in testing it out) to run

        blockdev --setra 2048 /dev/[sda/sda1/...]

before each benchmark to ensure fairness and simplicity of analysis, thanks.

Wu

Documentation/sysctl/vm.txt +37 -0
@@ -29,6 +29,8 @@ Currently, these files are in /proc/sys/
 - drop-caches
 - zone_reclaim_mode
 - zone_reclaim_interval
+- readahead_ratio
+- readahead_hit_rate
 
 ==============================================================
 
@@ -178,3 +180,38 @@ Time is set in seconds and set by defaul
 Reduce the interval if undesired off node allocations occur. However, too
 frequent scans will have a negative impact onoff node allocation performance.
 
+
+==============================================================
+
+readahead_ratio
+
+This limits readahead size to percent of the thrashing threshold.
+The thrashing threshold is dynamicly estimated from the _history_ read
+speed and system load, to deduce the _future_ readahead request size.
+
+Set it to a smaller value if you have not enough memory for all the
+concurrent readers, or the I/O loads fluctuate a lot. But if there's
+plenty of memory(>2MB per reader), a bigger value may help performance.
+
+readahead_ratio also selects the readahead logic:
+	VALUE	CODE PATH
+	-------------------------------------------
+	    0	disable readahead totally
+	  1-9	select the stock readahead logic
+	10-inf	select the adaptive readahead logic
+
+The default value is 50.  Reasonable values would be [50, 100].
+
+==============================================================
+
+readahead_hit_rate
+
+This is the max allowed value of (readahead-pages : accessed-pages).
+Useful only when (readahead_ratio >= 10). If the previous readahead
+request has bad hit rate, the kernel will be reluctant to do the next
+readahead.
+
+Larger values help catch more sparse access patterns. Be aware that
+readahead of the sparse patterns sacrifices memory for speed.
+
+The default value is 1.  It is recommended to keep the value below 16.
block/ll_rw_blk.c +31 -3
@@ -249,9 +249,6 @@ void blk_queue_make_request(request_queu
 	blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
 	blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
 	q->make_request_fn = mfn;
-	q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
-	q->backing_dev_info.state = 0;
-	q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
 	blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
 	blk_queue_hardsect_size(q, 512);
 	blk_queue_dma_alignment(q, 511);
@@ -1847,6 +1844,7 @@ request_queue_t *blk_alloc_queue_node(gf
 	q->kobj.ktype = &queue_ktype;
 	kobject_init(&q->kobj);
 
+	q->backing_dev_info = default_backing_dev_info;
 	q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
 	q->backing_dev_info.unplug_io_data = q;
 
@@ -3764,6 +3762,29 @@ queue_ra_store(struct request_queue *q, 
 	return ret;
 }
 
+static ssize_t queue_initial_ra_show(struct request_queue *q, char *page)
+{
+	int kb = q->backing_dev_info.ra_pages0 << (PAGE_CACHE_SHIFT - 10);
+
+	return queue_var_show(kb, (page));
+}
+
+static ssize_t
+queue_initial_ra_store(struct request_queue *q, const char *page, size_t count)
+{
+	unsigned long kb, ra;
+	ssize_t ret = queue_var_store(&kb, page, count);
+
+	ra = kb >> (PAGE_CACHE_SHIFT - 10);
+	q->backing_dev_info.ra_pages0 = ra;
+
+	ra = kb * 1024;
+	if (q->backing_dev_info.ra_expect_bytes > ra)
+		q->backing_dev_info.ra_expect_bytes = ra;
+
+	return ret;
+}
+
 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
 {
 	int max_sectors_kb = q->max_sectors >> 1;
@@ -3821,6 +3842,12 @@ static struct queue_sysfs_entry queue_ra
 	.store = queue_ra_store,
 };
 
+static struct queue_sysfs_entry queue_initial_ra_entry = {
+	.attr = {.name = "initial_ra_kb", .mode = S_IRUGO | S_IWUSR },
+	.show = queue_initial_ra_show,
+	.store = queue_initial_ra_store,
+};
+
 static struct queue_sysfs_entry queue_max_sectors_entry = {
 	.attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
 	.show = queue_max_sectors_show,
@@ -3841,6 +3868,7 @@ static struct queue_sysfs_entry queue_io
 static struct attribute *default_attrs[] = {
 	&queue_requests_entry.attr,
 	&queue_ra_entry.attr,
+	&queue_initial_ra_entry.attr,
 	&queue_max_hw_sectors_entry.attr,
 	&queue_max_sectors_entry.attr,
 	&queue_iosched_entry.attr,
drivers/block/loop.c +6 -0
@@ -779,6 +779,12 @@ static int loop_set_fd(struct loop_devic
 	mapping = file->f_mapping;
 	inode = mapping->host;
 
+	/*
+	 * The upper layer should already do proper look-ahead,
+	 * one more look-ahead here only ruins the cache hit rate.
+	 */
+	file->f_ra.flags |= RA_FLAG_NO_LOOKAHEAD;
+
 	if (!(file->f_mode & FMODE_WRITE))
 		lo_flags |= LO_FLAGS_READ_ONLY;
 
fs/file_table.c +7 -0
@@ -12,6 +12,7 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/smp_lock.h>
+#include <linux/mm.h>
 #include <linux/fs.h>
 #include <linux/security.h>
 #include <linux/eventpoll.h>
@@ -160,6 +161,12 @@ void fastcall __fput(struct file *file)
 	might_sleep();
 
 	fsnotify_close(file);
+
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+	if (file->f_ra.flags & RA_FLAG_EOF)
+		readahead_close(file);
+#endif
+
 	/*
 	 * The function eventpoll_release() should be the first called
 	 * in the file cleanup chain.
fs/mpage.c +3 -1
@@ -407,8 +407,10 @@ mpage_readpages(struct address_space *ma
 					&last_block_in_bio, &map_bh,
 					&first_logical_block,
 					get_block);
-			if (!pagevec_add(&lru_pvec, page))
+			if (!pagevec_add(&lru_pvec, page)) {
+				cond_resched();
 				__pagevec_lru_add(&lru_pvec);
+			}
 		} else {
 			page_cache_release(page);
 		}
fs/nfsd/vfs.c +4 -1
@@ -829,7 +829,10 @@ nfsd_vfs_read(struct svc_rqst *rqstp, st
 #endif
 
 	/* Get readahead parameters */
-	ra = nfsd_get_raparms(inode->i_sb->s_dev, inode->i_ino);
+	if (prefer_adaptive_readahead())
+		ra = NULL;
+	else
+		ra = nfsd_get_raparms(inode->i_sb->s_dev, inode->i_ino);
 
 	if (ra && ra->p_set)
 		file->f_ra = ra->p_ra;
include/linux/backing-dev.h +3 -0
@@ -24,6 +24,9 @@ typedef int (congested_fn)(void *, int);
 
 struct backing_dev_info {
 	unsigned long ra_pages;	/* max readahead in PAGE_CACHE_SIZE units */
+	unsigned long ra_pages0; /* recommended readahead on start of file */
+	unsigned long ra_expect_bytes;	/* expected read size on start of file */
+	unsigned long ra_thrash_bytes;	/* thrashing threshold */
 	unsigned long state;	/* Always use atomic bitops on this */
 	unsigned int capabilities; /* Device capabilities */
 	congested_fn *congested_fn; /* Function pointer if device is md/dm */
include/linux/fs.h +64 -10

85 lines not on this page · raw message

include/linux/mm.h +31 -0
@@ -955,7 +955,11 @@ extern int filemap_populate(struct vm_ar
 int write_one_page(struct page *page, int wait);
 
 /* readahead.c */
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+#define VM_MAX_READAHEAD	1024	/* kbytes */
+#else
 #define VM_MAX_READAHEAD	128	/* kbytes */
+#endif
 #define VM_MIN_READAHEAD	16	/* kbytes (includes current page) */
 #define VM_MAX_CACHE_HIT    	256	/* max pages in a row in cache before
 					 * turning readahead off */
@@ -972,6 +976,33 @@ unsigned long page_cache_readahead(struc
 void handle_ra_miss(struct address_space *mapping, 
 		    struct file_ra_state *ra, pgoff_t offset);
 unsigned long max_sane_readahead(unsigned long nr);
+void readahead_close(struct file *file);
+unsigned long
+page_cache_readahead_adaptive(struct address_space *mapping,
+			struct file_ra_state *ra, struct file *filp,
+			struct page *prev_page, struct page *page,
+			pgoff_t first_index, pgoff_t index, pgoff_t last_index);
+void readahead_cache_hit(struct file_ra_state *ra, struct page *page);
+
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+extern int readahead_ratio;
+#else
+#define readahead_ratio 1
+#endif /* CONFIG_ADAPTIVE_READAHEAD */
+
+static inline int prefer_adaptive_readahead(void)
+{
+	return readahead_ratio >= 10;
+}
+
+DECLARE_PER_CPU(unsigned long, readahead_aging);
+static inline void inc_readahead_aging(void)
+{
+#ifdef CONFIG_READAHEAD_SMOOTH_AGING
+	if (prefer_adaptive_readahead())
+		per_cpu(readahead_aging, raw_smp_processor_id())++;
+#endif
+}
 
 /* Do stack extension */
 extern int expand_stack(struct vm_area_struct *vma, unsigned long address);
include/linux/mmzone.h +6 -0
@@ -162,6 +162,11 @@ struct zone {
 	unsigned long		pages_scanned;	   /* since last reclaim */
 	int			all_unreclaimable; /* All pages pinned */
 
+	/* The accumulated number of activities that may cause page aging,
+	 * that is, make some pages closer to the tail of inactive_list.
+	 */
+	unsigned long 		aging_total;
+
 	/* A count of how many reclaimers are scanning this zone */
 	atomic_t		reclaim_in_progress;
 
@@ -328,6 +333,7 @@ void __get_zone_counts(unsigned long *ac
 			unsigned long *free, struct pglist_data *pgdat);
 void get_zone_counts(unsigned long *active, unsigned long *inactive,
 			unsigned long *free);
+unsigned long nr_free_inactive_pages_node(int nid);
 void build_all_zonelists(void);
 void wakeup_kswapd(struct zone *zone, int order);
 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
include/linux/page-flags.h +5 -0
@@ -89,6 +89,7 @@
 #define PG_buddy		19	/* Page is free, on buddy lists */
 
 #define PG_uncached		20	/* Page has been mapped as uncached */
+#define PG_readahead		21	/* Reminder to do readahead */
 
 /*
  * Global page accounting.  One instance per CPU.  Only unsigned longs are
@@ -360,6 +361,10 @@ extern void __mod_page_state_offset(unsi
 #define SetPageUncached(page)	set_bit(PG_uncached, &(page)->flags)
 #define ClearPageUncached(page)	clear_bit(PG_uncached, &(page)->flags)
 
+#define PageReadahead(page)	test_bit(PG_readahead, &(page)->flags)
+#define SetPageReadahead(page)	set_bit(PG_readahead, &(page)->flags)
+#define TestClearPageReadahead(page) test_and_clear_bit(PG_readahead, &(page)->flags)
+
 struct page;	/* forward declaration */
 
 int test_clear_page_dirty(struct page *page);
include/linux/pagemap.h +2 -0
@@ -68,6 +68,8 @@ static inline struct page *page_cache_al
 
 typedef int filler_t(void *, struct page *);
 
+extern int __probe_page(struct address_space *mapping, pgoff_t offset);
+extern int probe_page(struct address_space *mapping, pgoff_t offset);
 extern struct page * find_get_page(struct address_space *mapping,
 				unsigned long index);
 extern struct page * find_lock_page(struct address_space *mapping,
include/linux/radix-tree.h +4 -0
@@ -54,6 +54,10 @@ void *radix_tree_delete(struct radix_tre
 unsigned int
 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
 			unsigned long first_index, unsigned int max_items);
+unsigned long radix_tree_scan_hole_backward(struct radix_tree_root *root,
+				unsigned long index, unsigned long max_scan);
+unsigned long radix_tree_scan_hole(struct radix_tree_root *root,
+				unsigned long index, unsigned long max_scan);
 int radix_tree_preload(gfp_t gfp_mask);
 void radix_tree_init(void);
 void *radix_tree_tag_set(struct radix_tree_root *root,
include/linux/sysctl.h +2 -0
@@ -186,6 +186,8 @@ enum
 	VM_PERCPU_PAGELIST_FRACTION=30,/* int: fraction of pages in each percpu_pagelist */
 	VM_ZONE_RECLAIM_MODE=31, /* reclaim local zone memory before going off node */
 	VM_ZONE_RECLAIM_INTERVAL=32, /* time period to wait after reclaim failure */
+	VM_READAHEAD_RATIO=33,	/* percent of read-ahead size to thrashing-threshold */
+	VM_READAHEAD_HIT_RATE=34, /* one accessed page legitimizes so many read-ahead pages */
 };
 
 
include/linux/writeback.h +6 -0
@@ -85,6 +85,12 @@ void laptop_io_completion(void);
 void laptop_sync_completion(void);
 void throttle_vm_writeout(void);
 
+extern struct timer_list laptop_mode_wb_timer;
+static inline int laptop_spinned_down(void)
+{
+	return !timer_pending(&laptop_mode_wb_timer);
+}
+
 /* These are exported to sysctl. */
 extern int dirty_background_ratio;
 extern int vm_dirty_ratio;
kernel/sysctl.c +28 -0
@@ -73,6 +73,12 @@ extern int pid_max_min, pid_max_max;
 extern int sysctl_drop_caches;
 extern int percpu_pagelist_fraction;
 
+#if defined(CONFIG_ADAPTIVE_READAHEAD)
+extern int readahead_ratio;
+extern int readahead_hit_rate;
+static int one = 1;
+#endif
+
 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 int unknown_nmi_panic;
 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
@@ -915,6 +921,28 @@ static ctl_table vm_table[] = {
 		.strategy	= &sysctl_jiffies,
 	},
 #endif
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+	{
+		.ctl_name	= VM_READAHEAD_RATIO,
+		.procname	= "readahead_ratio",
+		.data		= &readahead_ratio,
+		.maxlen		= sizeof(readahead_ratio),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &zero,
+	},
+	{
+		.ctl_name	= VM_READAHEAD_HIT_RATE,
+		.procname	= "readahead_hit_rate",
+		.data		= &readahead_hit_rate,
+		.maxlen		= sizeof(readahead_hit_rate),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &one,
+	},
+#endif
 	{ .ctl_name = 0 }
 };
 
lib/radix-tree.c +71 -0

77 lines not on this page · raw message

mm/Kconfig +62 -0

65 lines not on this page · raw message

mm/filemap.c +102 -10

239 lines not on this page · raw message

mm/page-writeback.c +1 -1
@@ -376,7 +376,7 @@ static void wb_timer_fn(unsigned long un
 static void laptop_timer_fn(unsigned long unused);
 
 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
-static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
+DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 
 /*
  * Periodic writeback of "old" data.
mm/page_alloc.c +17 -1
@@ -543,7 +543,7 @@ static int prep_new_page(struct page *pa
 	if (PageReserved(page))
 		return 1;
 
-	page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
+	page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
 			1 << PG_referenced | 1 << PG_arch_1 |
 			1 << PG_checked | 1 << PG_mappedtodisk);
 	set_page_private(page, 0);
@@ -1361,6 +1361,22 @@ void get_zone_counts(unsigned long *acti
 	}
 }
 
+/*
+ * The node's effective length of inactive_list(s).
+ */
+unsigned long nr_free_inactive_pages_node(int nid)
+{
+	unsigned int i;
+	unsigned long sum = 0;
+	struct zone *zones = NODE_DATA(nid)->node_zones;
+
+	for (i = 0; i < MAX_NR_ZONES; i++)
+		sum += zones[i].nr_inactive +
+			zones[i].free_pages - zones[i].pages_low;
+
+	return sum;
+}
+
 void si_meminfo(struct sysinfo *val)
 {
 	val->totalram = totalram_pages;
mm/readahead.c +1601 -10

1,709 lines not on this page · raw message

mm/swap.c +2 -0
@@ -127,6 +127,8 @@ void fastcall mark_page_accessed(struct 
 		ClearPageReferenced(page);
 	} else if (!PageReferenced(page)) {
 		SetPageReferenced(page);
+		if (PageLRU(page))
+			inc_readahead_aging();
 	}
 }
 
mm/vmscan.c +4 -0
@@ -439,6 +439,9 @@ static unsigned long shrink_page_list(st
 		if (PageWriteback(page))
 			goto keep_locked;
 
+		if (!PageReferenced(page))
+			inc_readahead_aging();
+
 		referenced = page_referenced(page, 1);
 		/* In active use or really unfreeable?  Activate it. */
 		if (referenced && page_mapping_inuse(page))
@@ -637,6 +640,7 @@ static unsigned long shrink_inactive_lis
 					     &page_list, &nr_scan);
 		zone->nr_inactive -= nr_taken;
 		zone->pages_scanned += nr_scan;
+		zone->aging_total += nr_scan;
 		spin_unlock_irq(&zone->lru_lock);
 
 		nr_scanned += nr_scan;
author
#21 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 5 lines · On Tue, 6 Jun 2006 10:26:06 +0800 Fengguang Wu wrote:
Quoted 2 lines · Voluspa

Patch:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-git-updated-june-04-2006.patch

It seems that the patch has some problem:

[...]

Quoted 2 lines

The above statements was displaced, rendering the if() clause to fail all the time. That defeats the small file optimization, for ra_thrash_bytes will remain small.

Which rendered all my testing invalid. Nice... It came about with the update-01to04of04 and must have elicited a "fuzz" that I neglected to check.

Sorry to have caused you grief and extra work, Wu. I can only point towards the _Caveat and preemptive Mea Culpa_.

Quoted 6 lines

Voluspa, I attached an updated patch, including two more performance tunings. Please take it when you find time to do more benchmarks, thanks.

I'd suggest that you(and other kind people interested in testing it out) to run

        blockdev --setra 2048 /dev/[sda/sda1/...]

before each benchmark to ensure fairness and simplicity of analysis, thanks.

I'm in the process of writing up a new report after having tested for ca 6 hours straight (repenting mood). Will post in the original thread:

Adaptive Readahead V14 - statistics question...
http://marc.theaimsgroup.com/?t=114893205000004&r=1&w=2

The subject has a 'higher profile' than this one and might pull in future testers.

Note to Andrew; Revised Conclusion: On _this_ machine, with _these_ operations, Adaptive Readahead in its current incarnation and default settings is a slight _loss_. However, if the readahead size is lowered from 2048 to 256, it becomes a slight _gain_ or at least stays in parity with normal readahead.

Mvh
Mats Johannesson
--

#22 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 10 lines · Voluspa
Quoted 5 lines · On Tue, 6 Jun 2006 10:26:06 +0800 Fengguang Wu wrote:
Quoted 2 lines · Voluspa

Patch:
http://web.comhem.se/~u46139355/storetmp/adaptive-readahead-v14-linux-2.6.17-rc5-git-updated-june-04-2006.patch

It seems that the patch has some problem:

[...]

Quoted 2 lines

The above statements was displaced, rendering the if() clause to fail all the time. That defeats the small file optimization, for ra_thrash_bytes will remain small.

Quoted 6 lines

Which rendered all my testing invalid. Nice... It came about with the update-01to04of04 and must have elicited a "fuzz" that I neglected to check.

Sorry to have caused you grief and extra work, Wu. I can only point towards the _Caveat and preemptive Mea Culpa_.

Not bad ;-)
The stresses imposed forced me to think hard about the overheads the adaptive readahead introduced. And also some areas that the stock readahead has been good at.

me too, have some performance numbers, to be posted on the preferred thread.

Thanks,
Wu

#23 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

In reply to Voluspa

Voluspa,

Just updated the patch a little more, as attached :)

Would you test two simple commands by the way?

time (find / >/dev/null)
(repeat and drop the first result)

dd if=/dev/zero of=sparse bs=1M seek=5000 count=1

time cp sparse /dev/null
(repeat and drop the first result)

These commands measure the pure cpu(or software) overhead of the readahead logics, without ever hitting the physical device. I have no 64bit cpu numbers for them yet ;)

Thanks,
Wu

Documentation/sysctl/vm.txt +37 -0
@@ -29,6 +29,8 @@ Currently, these files are in /proc/sys/
 - drop-caches
 - zone_reclaim_mode
 - zone_reclaim_interval
+- readahead_ratio
+- readahead_hit_rate
 
 ==============================================================
 
@@ -178,3 +180,38 @@ Time is set in seconds and set by defaul
 Reduce the interval if undesired off node allocations occur. However, too
 frequent scans will have a negative impact onoff node allocation performance.
 
+
+==============================================================
+
+readahead_ratio
+
+This limits readahead size to percent of the thrashing threshold.
+The thrashing threshold is dynamicly estimated from the _history_ read
+speed and system load, to deduce the _future_ readahead request size.
+
+Set it to a smaller value if you have not enough memory for all the
+concurrent readers, or the I/O loads fluctuate a lot. But if there's
+plenty of memory(>2MB per reader), a bigger value may help performance.
+
+readahead_ratio also selects the readahead logic:
+	VALUE	CODE PATH
+	-------------------------------------------
+	    0	disable readahead totally
+	  1-9	select the stock readahead logic
+	10-inf	select the adaptive readahead logic
+
+The default value is 50.  Reasonable values would be [50, 100].
+
+==============================================================
+
+readahead_hit_rate
+
+This is the max allowed value of (readahead-pages : accessed-pages).
+Useful only when (readahead_ratio >= 10). If the previous readahead
+request has bad hit rate, the kernel will be reluctant to do the next
+readahead.
+
+Larger values help catch more sparse access patterns. Be aware that
+readahead of the sparse patterns sacrifices memory for speed.
+
+The default value is 1.  It is recommended to keep the value below 16.
block/ll_rw_blk.c +31 -3
@@ -249,9 +249,6 @@ void blk_queue_make_request(request_queu
 	blk_queue_max_phys_segments(q, MAX_PHYS_SEGMENTS);
 	blk_queue_max_hw_segments(q, MAX_HW_SEGMENTS);
 	q->make_request_fn = mfn;
-	q->backing_dev_info.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
-	q->backing_dev_info.state = 0;
-	q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
 	blk_queue_max_sectors(q, SAFE_MAX_SECTORS);
 	blk_queue_hardsect_size(q, 512);
 	blk_queue_dma_alignment(q, 511);
@@ -1847,6 +1844,7 @@ request_queue_t *blk_alloc_queue_node(gf
 	q->kobj.ktype = &queue_ktype;
 	kobject_init(&q->kobj);
 
+	q->backing_dev_info = default_backing_dev_info;
 	q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
 	q->backing_dev_info.unplug_io_data = q;
 
@@ -3764,6 +3762,29 @@ queue_ra_store(struct request_queue *q, 
 	return ret;
 }
 
+static ssize_t queue_initial_ra_show(struct request_queue *q, char *page)
+{
+	int kb = q->backing_dev_info.ra_pages0 << (PAGE_CACHE_SHIFT - 10);
+
+	return queue_var_show(kb, (page));
+}
+
+static ssize_t
+queue_initial_ra_store(struct request_queue *q, const char *page, size_t count)
+{
+	unsigned long kb, ra;
+	ssize_t ret = queue_var_store(&kb, page, count);
+
+	ra = kb >> (PAGE_CACHE_SHIFT - 10);
+	q->backing_dev_info.ra_pages0 = ra;
+
+	ra = kb * 1024;
+	if (q->backing_dev_info.ra_expect_bytes > ra)
+		q->backing_dev_info.ra_expect_bytes = ra;
+
+	return ret;
+}
+
 static ssize_t queue_max_sectors_show(struct request_queue *q, char *page)
 {
 	int max_sectors_kb = q->max_sectors >> 1;
@@ -3821,6 +3842,12 @@ static struct queue_sysfs_entry queue_ra
 	.store = queue_ra_store,
 };
 
+static struct queue_sysfs_entry queue_initial_ra_entry = {
+	.attr = {.name = "initial_ra_kb", .mode = S_IRUGO | S_IWUSR },
+	.show = queue_initial_ra_show,
+	.store = queue_initial_ra_store,
+};
+
 static struct queue_sysfs_entry queue_max_sectors_entry = {
 	.attr = {.name = "max_sectors_kb", .mode = S_IRUGO | S_IWUSR },
 	.show = queue_max_sectors_show,
@@ -3841,6 +3868,7 @@ static struct queue_sysfs_entry queue_io
 static struct attribute *default_attrs[] = {
 	&queue_requests_entry.attr,
 	&queue_ra_entry.attr,
+	&queue_initial_ra_entry.attr,
 	&queue_max_hw_sectors_entry.attr,
 	&queue_max_sectors_entry.attr,
 	&queue_iosched_entry.attr,
drivers/block/loop.c +6 -0
@@ -779,6 +779,12 @@ static int loop_set_fd(struct loop_devic
 	mapping = file->f_mapping;
 	inode = mapping->host;
 
+	/*
+	 * The upper layer should already do proper look-ahead,
+	 * one more look-ahead here only ruins the cache hit rate.
+	 */
+	file->f_ra.flags |= RA_FLAG_NO_LOOKAHEAD;
+
 	if (!(file->f_mode & FMODE_WRITE))
 		lo_flags |= LO_FLAGS_READ_ONLY;
 
fs/file_table.c +7 -0
@@ -12,6 +12,7 @@
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/smp_lock.h>
+#include <linux/mm.h>
 #include <linux/fs.h>
 #include <linux/security.h>
 #include <linux/eventpoll.h>
@@ -160,6 +161,12 @@ void fastcall __fput(struct file *file)
 	might_sleep();
 
 	fsnotify_close(file);
+
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+	if (file->f_ra.flags & RA_FLAG_EOF)
+		readahead_close(file);
+#endif
+
 	/*
 	 * The function eventpoll_release() should be the first called
 	 * in the file cleanup chain.
fs/mpage.c +3 -1
@@ -407,8 +407,10 @@ mpage_readpages(struct address_space *ma
 					&last_block_in_bio, &map_bh,
 					&first_logical_block,
 					get_block);
-			if (!pagevec_add(&lru_pvec, page))
+			if (!pagevec_add(&lru_pvec, page)) {
+				cond_resched();
 				__pagevec_lru_add(&lru_pvec);
+			}
 		} else {
 			page_cache_release(page);
 		}
fs/nfsd/vfs.c +4 -1
@@ -829,7 +829,10 @@ nfsd_vfs_read(struct svc_rqst *rqstp, st
 #endif
 
 	/* Get readahead parameters */
-	ra = nfsd_get_raparms(inode->i_sb->s_dev, inode->i_ino);
+	if (prefer_adaptive_readahead())
+		ra = NULL;
+	else
+		ra = nfsd_get_raparms(inode->i_sb->s_dev, inode->i_ino);
 
 	if (ra && ra->p_set)
 		file->f_ra = ra->p_ra;
include/linux/backing-dev.h +3 -0
@@ -24,6 +24,9 @@ typedef int (congested_fn)(void *, int);
 
 struct backing_dev_info {
 	unsigned long ra_pages;	/* max readahead in PAGE_CACHE_SIZE units */
+	unsigned long ra_pages0; /* recommended readahead on start of file */
+	unsigned long ra_expect_bytes;	/* expected read size on start of file */
+	unsigned long ra_thrash_bytes;	/* thrashing threshold */
 	unsigned long state;	/* Always use atomic bitops on this */
 	unsigned int capabilities; /* Device capabilities */
 	congested_fn *congested_fn; /* Function pointer if device is md/dm */
include/linux/fs.h +64 -10

85 lines not on this page · raw message

include/linux/mm.h +31 -0
@@ -955,7 +955,11 @@ extern int filemap_populate(struct vm_ar
 int write_one_page(struct page *page, int wait);
 
 /* readahead.c */
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+#define VM_MAX_READAHEAD	1024	/* kbytes */
+#else
 #define VM_MAX_READAHEAD	128	/* kbytes */
+#endif
 #define VM_MIN_READAHEAD	16	/* kbytes (includes current page) */
 #define VM_MAX_CACHE_HIT    	256	/* max pages in a row in cache before
 					 * turning readahead off */
@@ -972,6 +976,33 @@ unsigned long page_cache_readahead(struc
 void handle_ra_miss(struct address_space *mapping, 
 		    struct file_ra_state *ra, pgoff_t offset);
 unsigned long max_sane_readahead(unsigned long nr);
+void readahead_close(struct file *file);
+unsigned long
+page_cache_readahead_adaptive(struct address_space *mapping,
+			struct file_ra_state *ra, struct file *filp,
+			struct page *prev_page, struct page *page,
+			pgoff_t first_index, pgoff_t index, pgoff_t last_index);
+void readahead_cache_hit(struct file_ra_state *ra, struct page *page);
+
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+extern int readahead_ratio;
+#else
+#define readahead_ratio 1
+#endif /* CONFIG_ADAPTIVE_READAHEAD */
+
+static inline int prefer_adaptive_readahead(void)
+{
+	return readahead_ratio >= 10;
+}
+
+DECLARE_PER_CPU(unsigned long, readahead_aging);
+static inline void inc_readahead_aging(void)
+{
+#ifdef CONFIG_READAHEAD_SMOOTH_AGING
+	if (prefer_adaptive_readahead())
+		per_cpu(readahead_aging, raw_smp_processor_id())++;
+#endif
+}
 
 /* Do stack extension */
 extern int expand_stack(struct vm_area_struct *vma, unsigned long address);
include/linux/mmzone.h +6 -0
@@ -162,6 +162,11 @@ struct zone {
 	unsigned long		pages_scanned;	   /* since last reclaim */
 	int			all_unreclaimable; /* All pages pinned */
 
+	/* The accumulated number of activities that may cause page aging,
+	 * that is, make some pages closer to the tail of inactive_list.
+	 */
+	unsigned long 		aging_total;
+
 	/* A count of how many reclaimers are scanning this zone */
 	atomic_t		reclaim_in_progress;
 
@@ -328,6 +333,7 @@ void __get_zone_counts(unsigned long *ac
 			unsigned long *free, struct pglist_data *pgdat);
 void get_zone_counts(unsigned long *active, unsigned long *inactive,
 			unsigned long *free);
+unsigned long nr_free_inactive_pages_node(int nid);
 void build_all_zonelists(void);
 void wakeup_kswapd(struct zone *zone, int order);
 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
include/linux/page-flags.h +5 -0
@@ -89,6 +89,7 @@
 #define PG_buddy		19	/* Page is free, on buddy lists */
 
 #define PG_uncached		20	/* Page has been mapped as uncached */
+#define PG_readahead		21	/* Reminder to do readahead */
 
 /*
  * Global page accounting.  One instance per CPU.  Only unsigned longs are
@@ -360,6 +361,10 @@ extern void __mod_page_state_offset(unsi
 #define SetPageUncached(page)	set_bit(PG_uncached, &(page)->flags)
 #define ClearPageUncached(page)	clear_bit(PG_uncached, &(page)->flags)
 
+#define PageReadahead(page)	test_bit(PG_readahead, &(page)->flags)
+#define SetPageReadahead(page)	set_bit(PG_readahead, &(page)->flags)
+#define TestClearPageReadahead(page) test_and_clear_bit(PG_readahead, &(page)->flags)
+
 struct page;	/* forward declaration */
 
 int test_clear_page_dirty(struct page *page);
include/linux/pagemap.h +2 -0
@@ -68,6 +68,8 @@ static inline struct page *page_cache_al
 
 typedef int filler_t(void *, struct page *);
 
+extern int __probe_page(struct address_space *mapping, pgoff_t offset);
+extern int probe_page(struct address_space *mapping, pgoff_t offset);
 extern struct page * find_get_page(struct address_space *mapping,
 				unsigned long index);
 extern struct page * find_lock_page(struct address_space *mapping,
include/linux/radix-tree.h +4 -0
@@ -54,6 +54,10 @@ void *radix_tree_delete(struct radix_tre
 unsigned int
 radix_tree_gang_lookup(struct radix_tree_root *root, void **results,
 			unsigned long first_index, unsigned int max_items);
+unsigned long radix_tree_scan_hole_backward(struct radix_tree_root *root,
+				unsigned long index, unsigned long max_scan);
+unsigned long radix_tree_scan_hole(struct radix_tree_root *root,
+				unsigned long index, unsigned long max_scan);
 int radix_tree_preload(gfp_t gfp_mask);
 void radix_tree_init(void);
 void *radix_tree_tag_set(struct radix_tree_root *root,
include/linux/sysctl.h +2 -0
@@ -186,6 +186,8 @@ enum
 	VM_PERCPU_PAGELIST_FRACTION=30,/* int: fraction of pages in each percpu_pagelist */
 	VM_ZONE_RECLAIM_MODE=31, /* reclaim local zone memory before going off node */
 	VM_ZONE_RECLAIM_INTERVAL=32, /* time period to wait after reclaim failure */
+	VM_READAHEAD_RATIO=33,	/* percent of read-ahead size to thrashing-threshold */
+	VM_READAHEAD_HIT_RATE=34, /* one accessed page legitimizes so many read-ahead pages */
 };
 
 
include/linux/writeback.h +6 -0
@@ -85,6 +85,12 @@ void laptop_io_completion(void);
 void laptop_sync_completion(void);
 void throttle_vm_writeout(void);
 
+extern struct timer_list laptop_mode_wb_timer;
+static inline int laptop_spinned_down(void)
+{
+	return !timer_pending(&laptop_mode_wb_timer);
+}
+
 /* These are exported to sysctl. */
 extern int dirty_background_ratio;
 extern int vm_dirty_ratio;
kernel/sysctl.c +28 -0
@@ -73,6 +73,12 @@ extern int pid_max_min, pid_max_max;
 extern int sysctl_drop_caches;
 extern int percpu_pagelist_fraction;
 
+#if defined(CONFIG_ADAPTIVE_READAHEAD)
+extern int readahead_ratio;
+extern int readahead_hit_rate;
+static int one = 1;
+#endif
+
 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 int unknown_nmi_panic;
 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
@@ -915,6 +921,28 @@ static ctl_table vm_table[] = {
 		.strategy	= &sysctl_jiffies,
 	},
 #endif
+#ifdef CONFIG_ADAPTIVE_READAHEAD
+	{
+		.ctl_name	= VM_READAHEAD_RATIO,
+		.procname	= "readahead_ratio",
+		.data		= &readahead_ratio,
+		.maxlen		= sizeof(readahead_ratio),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &zero,
+	},
+	{
+		.ctl_name	= VM_READAHEAD_HIT_RATE,
+		.procname	= "readahead_hit_rate",
+		.data		= &readahead_hit_rate,
+		.maxlen		= sizeof(readahead_hit_rate),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &one,
+	},
+#endif
 	{ .ctl_name = 0 }
 };
 
lib/radix-tree.c +71 -0

77 lines not on this page · raw message

mm/Kconfig +62 -0

65 lines not on this page · raw message

mm/filemap.c +102 -10

239 lines not on this page · raw message

mm/page-writeback.c +1 -1
@@ -376,7 +376,7 @@ static void wb_timer_fn(unsigned long un
 static void laptop_timer_fn(unsigned long unused);
 
 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
-static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
+DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 
 /*
  * Periodic writeback of "old" data.
mm/page_alloc.c +17 -1
@@ -543,7 +543,7 @@ static int prep_new_page(struct page *pa
 	if (PageReserved(page))
 		return 1;
 
-	page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
+	page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
 			1 << PG_referenced | 1 << PG_arch_1 |
 			1 << PG_checked | 1 << PG_mappedtodisk);
 	set_page_private(page, 0);
@@ -1361,6 +1361,22 @@ void get_zone_counts(unsigned long *acti
 	}
 }
 
+/*
+ * The node's effective length of inactive_list(s).
+ */
+unsigned long nr_free_inactive_pages_node(int nid)
+{
+	unsigned int i;
+	unsigned long sum = 0;
+	struct zone *zones = NODE_DATA(nid)->node_zones;
+
+	for (i = 0; i < MAX_NR_ZONES; i++)
+		sum += zones[i].nr_inactive +
+			zones[i].free_pages - zones[i].pages_low;
+
+	return sum;
+}
+
 void si_meminfo(struct sysinfo *val)
 {
 	val->totalram = totalram_pages;
mm/readahead.c +1601 -10

1,709 lines not on this page · raw message

mm/swap.c +2 -0
@@ -127,6 +127,8 @@ void fastcall mark_page_accessed(struct 
 		ClearPageReferenced(page);
 	} else if (!PageReferenced(page)) {
 		SetPageReferenced(page);
+		if (PageLRU(page))
+			inc_readahead_aging();
 	}
 }
 
mm/vmscan.c +4 -0
@@ -439,6 +439,9 @@ static unsigned long shrink_page_list(st
 		if (PageWriteback(page))
 			goto keep_locked;
 
+		if (!PageReferenced(page))
+			inc_readahead_aging();
+
 		referenced = page_referenced(page, 1);
 		/* In active use or really unfreeable?  Activate it. */
 		if (referenced && page_mapping_inuse(page))
@@ -637,6 +640,7 @@ static unsigned long shrink_inactive_lis
 					     &page_list, &nr_scan);
 		zone->nr_inactive -= nr_taken;
 		zone->pages_scanned += nr_scan;
+		zone->aging_total += nr_scan;
 		spin_unlock_irq(&zone->lru_lock);
 
 		nr_scanned += nr_scan;
author
#24 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

On Tue, 6 Jun 2006 10:57:28 +0800 Fengguang Wu wrote:
[...]

Quoted 5 lines

Would you test two simple commands by the way?

time (find / >/dev/null)
(repeat and drop the first result)

Quoted 8 lines

dd if=/dev/zero of=sparse bs=1M seek=5000 count=1

time cp sparse /dev/null
(repeat and drop the first result)

These commands measure the pure cpu(or software) overhead of the readahead logics, without ever hitting the physical device. I have no 64bit cpu numbers for them yet ;)

The commands don't make sense (I'm in bash)... Also, I refuse to overwrite /dev/null ;-) I've done a test with "cat" but the first and second run of it are the same, so it does touch the disk, it seems:

root:sleipner:~# grep setra /etc/rc.d/rc.local
#/sbin/blockdev --setra 2048 /dev/hda
/sbin/blockdev --setra 256 /dev/hda
root:sleipner:~# uname -r
2.6.17-rc5-git10-ar2
root:sleipner:~# time find / >/dev/null
real    3m34.119s
user    0m1.118s
sys     0m5.844s
root:sleipner:~# time find / >/dev/null
real    0m2.949s
user    0m0.691s
sys     0m2.253s
root:sleipner:~# dd if=/dev/zero of=sparse bs=1M seek=5000 count=1
1+0 records in
1+0 records out
root:sleipner:~# time cat sparse >/dev/null
real    0m9.377s
user    0m0.140s
sys     0m8.715s
root:sleipner:~# time cat sparse >/dev/null
real    0m9.593s
user    0m0.138s
sys     0m8.680s

*******************************************************************

root:sleipner:~# grep setra /etc/rc.d/rc.local
/sbin/blockdev --setra 2048 /dev/hda
#/sbin/blockdev --setra 256 /dev/hda
root:sleipner:~# uname -r
2.6.17-rc5-git10-ar2
root:sleipner:~# time find / >/dev/null
real    3m33.037s
user    0m1.112s
sys     0m5.821s
root:sleipner:~# time find / >/dev/null
real    0m2.538s
user    0m0.622s
sys     0m1.911s
root:sleipner:~# dd if=/dev/zero of=sparse bs=1M seek=5000 count=1
1+0 records in
1+0 records out
root:sleipner:~# time cat sparse >/dev/null
real    0m11.279s
user    0m0.163s
sys     0m10.340s
root:sleipner:~# time cat sparse >/dev/null
real    0m11.268s
user    0m0.154s
sys     0m10.412s

Mvh
Mats Johannesson
--

#25 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

It's interesting that copying of sparse file is more efficient with small readahead size :) I get the same conclusion, though with smaller differences:

SUMMARY

              user       sys       cpu         total
ARA, 1M       0.15       6.28      82.80       7.73
ARA, 128k     0.14       6.09      85.60       7.26
STOCK, 128k   0.15       6.05      85.60       7.22

TEST CASE

wfg ~% ll work/sparse
-rw-r--r-- 1 wfg wfg 1.6G 2006-05-21 15:11 work/sparse wfg ~% free
total used free shared buffers cached
Mem: 501 496 5 0 5 191
-/+ buffers/cache: 300 201
Swap: 127 0 127

wfg ~% time cp work/sparse /dev/null

ARA, 1M

cp work/sparse /dev/null  0.15s user 6.35s system 80% cpu 8.125 total
cp work/sparse /dev/null  0.14s user 6.28s system 84% cpu 7.556 total
cp work/sparse /dev/null  0.14s user 6.25s system 82% cpu 7.744 total
cp work/sparse /dev/null  0.15s user 6.23s system 85% cpu 7.495 total
cp work/sparse /dev/null  0.16s user 6.30s system 83% cpu 7.719 total

ARA, 128k

cp work/sparse /dev/null  0.15s user 6.07s system 86% cpu 7.224 total
cp work/sparse /dev/null  0.15s user 6.05s system 84% cpu 7.334 total
cp work/sparse /dev/null  0.14s user 6.18s system 86% cpu 7.328 total
cp work/sparse /dev/null  0.13s user 6.11s system 86% cpu 7.217 total
cp work/sparse /dev/null  0.14s user 6.06s system 86% cpu 7.179 total

STOCK, 128k

cp work/sparse /dev/null  0.16s user 6.01s system 86% cpu 7.162 total
cp work/sparse /dev/null  0.14s user 6.10s system 86% cpu 7.222 total
cp work/sparse /dev/null  0.14s user 6.04s system 86% cpu 7.186 total
cp work/sparse /dev/null  0.15s user 6.02s system 85% cpu 7.210 total
cp work/sparse /dev/null  0.14s user 6.09s system 85% cpu 7.320 total

Wu

author
#26 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 2 lines · On Thu, 8 Jun 2006 16:13:52 +0800 Fengguang Wu wrote:

It's interesting that copying of sparse file is more efficient with small readahead size :) I get the same conclusion, though with smaller differences:

How on earth can you copy the file without overwriting the target /dev/null? As you saw, I could just "cat" the file. Size was:

root:sleipner:~# dd if=/dev/zero of=sparse bs=1M seek=5000 count=1
1+0 records in
1+0 records out
root:sleipner:~# ls -l sparse 
1040 -rw-r--r--  1 root root 5243928576 Jun  8 10:26 sparse

5.2 fake GBs...

Mvh
Mats Johannesson
--

#27 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 5 lines · Voluspa
Quoted 2 lines · On Thu, 8 Jun 2006 16:13:52 +0800 Fengguang Wu wrote:

It's interesting that copying of sparse file is more efficient with small readahead size :) I get the same conclusion, though with smaller differences:

How on earth can you copy the file without overwriting the target /dev/null?

Yes, it worked as expected. All the time.

I'm running zsh, though bash is also tested ok.

% cp --version
cp (GNU coreutils) 5.96
Copyright (C) 2006 Free Software Foundation, Inc.

Wu

author
#28 lore raw

Re: [PATCH] readahead: initial method - expected read size - fix fastcall

Quoted 7 lines · On Thu, 8 Jun 2006 16:50:55 +0800 Fengguang Wu wrote:
Quoted 5 lines · Voluspa
Quoted 2 lines · On Thu, 8 Jun 2006 16:13:52 +0800 Fengguang Wu wrote:

It's interesting that copying of sparse file is more efficient with small readahead size :) I get the same conclusion, though with smaller differences:

How on earth can you copy the file without overwriting the target /dev/null?

Quoted 7 lines

Yes, it worked as expected. All the time.

I'm running zsh, though bash is also tested ok.

% cp --version
cp (GNU coreutils) 5.96
Copyright (C) 2006 Free Software Foundation, Inc.

root:sleipner:~# dd if=/dev/zero of=sparse bs=1M seek=5000 count=1
1+0 records in
1+0 records out
root:sleipner:~# time cp sparse /dev/null
cp: overwrite `/dev/null'? n
root:sleipner:~# cp --version
cp (coreutils) 5.2.1
Written by Torbjorn Granlund, David MacKenzie, and Jim Meyering.

Copyright (C) 2004 Free Software Foundation, Inc.
[etc...]

Ok, guess it's time to build a more modern system. Was planning to do it anyway. I don't like just updating something like coreutils...

Mvh
Mats Johannesson
--

#29 lore raw

PATCH 1/1 readahead: call scheme - fix fastcall readahead_cache_hit()

In reply to Andrew Morton

Remove 'fastcall' directive for readahead_cache_hit().

Sorry, it also leads to unfavorable performance in the following micro benchmark on i386 with CONFIG_REGPARM=n:

Command:
time cp cold /dev/null

Summary:

              user     sys     cpu    total
no-fastcall   1.24    24.88    90.9   28.57
fastcall      1.16    25.69    91.5   29.23

Details:
without fastcall:

cp cold /dev/null  1.27s user 24.63s system 91% cpu 28.348 total
cp cold /dev/null  1.17s user 25.09s system 91% cpu 28.653 total
cp cold /dev/null  1.24s user 24.75s system 91% cpu 28.448 total
cp cold /dev/null  1.20s user 25.04s system 91% cpu 28.614 total
cp cold /dev/null  1.31s user 24.67s system 91% cpu 28.499 total
cp cold /dev/null  1.30s user 24.87s system 91% cpu 28.530 total
cp cold /dev/null  1.26s user 24.84s system 91% cpu 28.542 total
cp cold /dev/null  1.16s user 25.15s system 90% cpu 28.925 total

with fastcall:

cp cold /dev/null  1.16s user 26.39s system 91% cpu 30.061 total
cp cold /dev/null  1.25s user 26.53s system 91% cpu 30.378 total
cp cold /dev/null  1.10s user 25.32s system 92% cpu 28.679 total
cp cold /dev/null  1.15s user 25.20s system 91% cpu 28.747 total
cp cold /dev/null  1.19s user 25.38s system 92% cpu 28.841 total
cp cold /dev/null  1.11s user 25.75s system 92% cpu 29.126 total
cp cold /dev/null  1.17s user 25.49s system 91% cpu 29.042 total
cp cold /dev/null  1.17s user 25.49s system 92% cpu 28.970 total

---

include/linux/mm.h +1 -1
@@ -1074,7 +1074,7 @@ page_cache_readahead_adaptive(struct add
 			struct file_ra_state *ra, struct file *filp,
 			struct page *prev_page, struct page *page,
 			pgoff_t first_index, pgoff_t index, pgoff_t last_index);
-void fastcall readahead_cache_hit(struct file_ra_state *ra, struct page *page);
+void readahead_cache_hit(struct file_ra_state *ra, struct page *page);
 
 #ifdef CONFIG_ADAPTIVE_READAHEAD
 extern int readahead_ratio;
mm/readahead.c +1 -1
@@ -1916,7 +1916,7 @@ readit:
  * readahead_cache_hit() is the feedback route of the adaptive read-ahead
  * logic. It must be called on every access on the read-ahead pages.
  */
-void fastcall readahead_cache_hit(struct file_ra_state *ra, struct page *page)
+void readahead_cache_hit(struct file_ra_state *ra, struct page *page)
 {
 	if (PageActive(page) || PageReferenced(page))
 		return;