Official thread: "zen-sources" - Part 7

[AaronPPC]

So "frakkin' fast" isn't quantitative enough?

[darklegion]

Guys, again

Can somebody post some REAL numbers on your experience?
Not super-puper boost/increase/decrease, but something that can be compared.
Interbench (http://users.on.net/~ckolivas/interbench/) numbers are ok, I think.

Here you go:

2.6.30-zen5 (cfs)

Code: Select all

Using 3568151 loops per ms, running every load for 30 seconds
Benchmarking kernel 2.6.30-zen5-31380-ge7c0d6d at datestamp 200908271859

--- Benchmarking simulated cpu of Audio in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.004 +/- 0.00459    0.011             100            100
Video     0.011 +/- 0.122       2.97             100            100
X         0.003 +/- 0.00328    0.008             100            100
Burn      0.003 +/- 0.00346    0.008             100            100
Write     0.424 +/- 6.25         141            99.8           99.8
Read      0.016 +/- 0.177       4.32             100            100
Compile   0.076 +/- 0.684       8.83             100            100
Memload   0.018 +/- 0.114       2.78             100            100

--- Benchmarking simulated cpu of Video in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.003 +/- 0.00392    0.009             100            100
X         0.004 +/- 0.00588    0.077             100            100
Burn       6.21 +/- 11.8          50            98.7           87.2
Write     0.262 +/- 3.24        73.3            99.5             99
Read      0.061 +/- 0.574       16.8             100           99.9
Compile    7.73 +/- 16.4        77.9            93.6           76.7
Memload   0.043 +/- 0.549       16.7             100           99.9

--- Benchmarking simulated cpu of X in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None          0 +/- 0.000619   0.008             100            100
Video         0 +/- 0.000619   0.007             100            100
Burn       51.4 +/- 77.3         201            18.4           9.46
Write      6.45 +/- 21.7         175            67.2           61.8
Read       23.2 +/- 41.4         112              30           20.3
Compile    68.8 +/- 98.9         238            15.4           6.71
Memload       0 +/- 0.00239     0.03             100            100

--- Benchmarking simulated cpu of Gaming in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU
None       0.07 +/- 1.1           19            99.9
Video     0.656 +/- 5.67        65.5            99.3
X         0.127 +/- 0.615        3.8            99.9
Burn        155 +/- 158          227            39.3
Write      18.3 +/- 41.6         187            84.5
Read       96.8 +/- 102          149            50.8
Compile     244 +/- 250          355            29.1
Memload   0.001 +/- 0.0161     0.277             100

2.6.30-zen5-bfs

Code: Select all

Using 3568151 loops per ms, running every load for 30 seconds
Benchmarking kernel 2.6.30-zen5-bfs-31391-g719faef at datestamp 200908271827

--- Benchmarking simulated cpu of Audio in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.003 +/- 0.00389    0.009             100            100
Video     0.002 +/- 0.00289    0.007             100            100
X         0.003 +/- 0.00333    0.008             100            100
Burn      0.003 +/- 0.00314    0.007             100            100
Write     0.014 +/- 0.121       2.94             100            100
Read      0.008 +/- 0.0098     0.122             100            100
Compile    0.01 +/- 0.0135     0.188             100            100
Memload   0.013 +/- 0.0282     0.249             100            100

--- Benchmarking simulated cpu of Video in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.003 +/- 0.00369    0.008             100            100
X         0.003 +/- 0.00322    0.008             100            100
Burn       14.6 +/- 15.6        16.7             100           12.6
Write     0.032 +/- 0.56        16.7             100           99.9
Read      0.006 +/- 0.00748     0.13             100            100
Compile    15.9 +/- 18            50            83.6             14
Memload   0.008 +/- 0.00948    0.032             100            100

--- Benchmarking simulated cpu of X in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None          0 +/- 0.000574   0.008             100            100
Video      0.01 +/- 0.129          2             100           99.3
Burn       53.2 +/- 73.1         156            19.9           8.63
Write      1.06 +/- 7.97          99            92.5           90.4
Read      0.033 +/- 0.258          2            98.7           97.7
Compile    75.1 +/- 100          219            15.9           5.48
Memload   0.026 +/- 0.283          4            99.3           98.7

--- Benchmarking simulated cpu of Gaming in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU
None          0 +/- 0.00869    0.151             100
Video     0.016 +/- 0.166       2.32             100
X         0.005 +/- 0.0893      1.55             100
Burn        148 +/- 149          164            40.3
Write      5.33 +/- 15          96.1            94.9
Read       1.78 +/- 2.1         5.46            98.3
Compile     223 +/- 226          287            30.9
Memload   0.307 +/- 1.6         17.3            99.7

Specifications:
Intel e5200 @ 3.6ghz
4GB DDR800 ram
Gigabyte ep45-ds3
Nvidia GTX275

Gentoo ~AMD64

Benchmark was run in X, but with only ratpoison and a single urxvt opened.Would have been better if it were run in runlevel 1, but I couldn't be bothered running it again.

[ponciarello]

init 1

Code: Select all

Using 2183223 loops per ms, running every load for 30 seconds
Benchmarking kernel 2.6.30.5 at datestamp 200908271114

--- Benchmarking simulated cpu of Audio in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.051 +/- 0.0511     0.057             100            100
Video     0.005 +/- 0.00597    0.012             100            100
X         0.033 +/- 0.0371      0.07             100            100
Burn      0.005 +/- 0.00516    0.006             100            100
Write     0.241 +/- 2.53        50.2             100           99.8
Read      0.009 +/- 0.0099     0.082             100            100
Compile   0.008 +/- 0.0216     0.481             100            100
Memload   0.036 +/- 0.392       9.03             100            100

--- Benchmarking simulated cpu of Video in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.054 +/- 0.055      0.078             100            100
X         0.043 +/- 0.0456     0.079             100            100
Burn       9.51 +/- 14.9        69.4              99             77
Write     0.126 +/- 2.01        69.7            99.8           99.6
Read      0.009 +/- 0.0186     0.218             100            100
Compile    8.39 +/- 15.1        79.9            96.7           77.1
Memload   0.051 +/- 0.0562     0.597             100            100

--- Benchmarking simulated cpu of X in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None       0.07 +/- 0.436          3              98             96
Video     0.136 +/- 0.719          4            95.6           93.5
Burn       53.8 +/- 80.5         209            17.6           8.36
Write       0.4 +/- 3.61          45            92.6           90.2
Read      0.136 +/- 0.719          4            95.6           93.5
Compile    54.5 +/- 81.8         210              18           8.89
Memload   0.192 +/- 0.917          6            95.3           92.2

--- Benchmarking simulated cpu of Gaming in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU
None       2.33 +/- 2.34        2.87            97.7
Video      3.63 +/- 3.64        3.89            96.5
X          3.64 +/- 3.72        5.14            96.5
Burn        166 +/- 172          244            37.6
Write       3.9 +/- 8.32         109            96.2
Read       3.36 +/- 3.38         4.2            96.7
Compile     180 +/- 186          290            35.7
Memload    3.83 +/- 4.53        42.5            96.3

Code: Select all

Using 2183223 loops per ms, running every load for 30 seconds
Benchmarking kernel 2.6.30-zen5-bfs at datestamp 200908270843

--- Benchmarking simulated cpu of Audio in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.006 +/- 0.00665    0.012             100            100
Video     0.005 +/- 0.00582    0.013             100            100
X         0.005 +/- 0.00611    0.009             100            100
Burn      0.069 +/- 1.1         19.1             100            100
Write     0.008 +/- 0.00858    0.022             100            100
Read      0.008 +/- 0.00888    0.022             100            100
Compile   0.016 +/- 0.187       4.56             100            100
Memload    0.01 +/- 0.0112     0.034             100            100

--- Benchmarking simulated cpu of Video in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.006 +/- 0.00682    0.015             100            100
X         0.006 +/- 0.00622    0.014             100            100
Burn       15.5 +/- 16.1        16.7             100           6.97
Write     0.017 +/- 0.393       16.7             100           99.9
Read      0.007 +/- 0.00732    0.027             100            100
Compile    15.7 +/- 16.4          50            98.1           7.04
Memload   0.008 +/- 0.00914    0.064             100            100

--- Benchmarking simulated cpu of X in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None      0.013 +/- 0.141          2             100             99
Video     0.069 +/- 0.436          3              98             96
Burn       55.9 +/- 76.7         158            19.7           8.27
Write     0.166 +/- 1.59          24            96.2           95.1
Read      0.049 +/- 0.361          3              98           96.7
Compile    60.5 +/- 83.7         220            18.4           7.68
Memload   0.219 +/- 1.26           9            93.8           91.7

--- Benchmarking simulated cpu of Gaming in the presence of simulated ---
Load    Latency +/- SD (ms)  Max Latency   % Desired CPU
None      0.714 +/- 0.717       1.03            99.3
Video      2.44 +/- 2.45         3.1            97.6
X          2.68 +/- 2.91        4.47            97.4
Burn        161 +/- 162          178            38.3
Write      2.13 +/- 3.98          29            97.9
Read       1.88 +/- 1.9         2.66            98.2
Compile     175 +/- 177          228            36.3
Memload    3.85 +/- 4.9         15.9            96.3

lspci -v, /proc/cpuinfo, this is the laptop.

[MageSlayer]

ponciarello
Well, no significant effect is seen, afaiu.
Except sudden drop in Burn in the 2d test.
Compile phase also does not show any wonders (Are you sure your joy about "make -j4" was not subjective ? )

darklegion
Here we see some more interesting picture - Read/Write in the 3d test is definitely better.

BTW, does anybody of you uses http://bugzilla.kernel.org/show_bug.cgi?id=12309#c397 patch in your kernel?
Sometimes people report it beats any those cpu scheduler tweaks.
As you understand we need a fair comparison

Thanks for info, guys

[ponciarello]

you can do some test too, mageslayer

look well at test results, i can say difference is tangible for the user: for me (also I play a lot Enemy Territory) it is

[kernelOfTruth]

I can't understand why those guys more or less neglect desktop-related optimizations

This probably isn't the right thread to debate this, but I wouldn't say he is ignoring the desktop. It is just that the Linux world is bigger than the desktop. The Linux Foundation has corporate backers that who care about servers and routers and not the desktop. I don't mean to imply that there is some corporate conspiracy. My point is that Linus has more things to worry about and he seems very opposed to forking the kernel into "home" and "server" products.

you're right, afaik they've already discussed this at length on lkml

My Point is: since nowadays linux is gaining more and more part in mobile and other "desktop" devices why settle for using an scheduler with excellent performance on multi-cpu environments and an mediocre performance on oligo-cpu environments and not use what's working great on those devices (== using 2 different selectable schedulers for those different use-cases) ?

anyways this shall be discussed on lkml or some other basis

BTW, does anybody of you uses http://bugzilla.kernel.org/show_bug.cgi?id=12309#c397 patch in your kernel?
Sometimes people report it beats any those cpu scheduler tweaks.
As you understand we need a fair comparison

yes ! everyone who's using 2.6.30-zen4

it really seems to make a noticable (but not significant) positive difference

[tranquilcool]

for now doesn't play well with CONFIG_DEVTMPFS set.
and microcode seg faults.

[kernelOfTruth]

for now doesn't play well with CONFIG_DEVTMPFS set.
and microcode seg faults.

for me it wouldn't even launch metacity, gnome, xfce4, midori (gtk-webkit based), chromium wouldn't launch any websites, ...

so I'll stay away from it until it has got some more testing (maybe it doesn't like hardening / SSP ? )

[ponciarello]

xfce4 works fine here

[cheater1034]

the cfq drain async i/o patch has been in zen for quite some time now (2.6.30 and 2.6.31)

From the charts i've seen (and the charts i've posted a page back, mageslayer) - there is a positive difference in numerous aspects.

And remember, interbench is an interactivity test, try the other benchmarks for thoroughput and responsiveness:
http://users.on.net/~ckolivas/

[cheater1034]

and btw, the origin/sched-bfs branch can be merged into master so you can try bfs on 2.6.31

[Waninkoko]

and btw, the origin/sched-bfs branch can be merged into master so you can try bfs on 2.6.31

But it's not recommened to use it yet as there are some remaining bugs I have to fix first.

[cheater1034]

master-2.6.30-bfs will soon be deleted, due to me splitting up bfs and cfs via Kconfig (like in the old days)

All the includes and declarations and a couple functions will be universal in sched.c, rest will be in either sched_cfs.c or sched_bfs.c

*edit*
use master-2.6.30 now please, bfs is in general setup to enable it

[darklegion]

for now doesn't play well with CONFIG_DEVTMPFS set.
and microcode seg faults.

for me it wouldn't even launch metacity, gnome, xfce4, midori (gtk-webkit based), chromium wouldn't launch any websites, ...

so I'll stay away from it until it has got some more testing (maybe it doesn't like hardening / SSP ? )

If you tried it when it was bfs-016, that would explain it.It was pretty much unusable for me.But bfs-018 works great.

[kernelOfTruth]

for now doesn't play well with CONFIG_DEVTMPFS set.
and microcode seg faults.

for me it wouldn't even launch metacity, gnome, xfce4, midori (gtk-webkit based), chromium wouldn't launch any websites, ...

so I'll stay away from it until it has got some more testing (maybe it doesn't like hardening / SSP ? )

If you tried it when it was bfs-016, that would explain it.It was pretty much unusable for me.But bfs-018 works great.

strange then bfs-019 got broken again ?

afaik it was bfs-019 that I was using

now let's see if 102 works better (100 <-- should be "stable" or at least usable now according to the version number)

[tranquilcool]

102 doesn't start hald. so keyboard and
mouse don't work. 018 only breaks microcode.

[darklegion]

strange then bfs-019 got broken again ?

afaik it was bfs-019 that I was using

now let's see if 102 works better (100 <-- should be "stable" or at least usable now according to the version number)

Yeah, it seems to be pretty volatile at the moment.I pulled in the updated master-2.6.30 with bfs merged in, and it often hangs while booting the kernel.This was yesterday, so I've just upgraded again today, and will see if it works well.If not, I'll just revert to master-2.6.30-bfs which doesn't have any issues on my system.

[bollucks]

Trying bfs103 on unpatched 2.6.30 boots and works flawlessly, but this patch does not. Broken merge?

[cheater1034]

Trying bfs103 on unpatched 2.6.30 boots and works flawlessly, but this patch does not. Broken merge?

bfs 103 works up to 2.6.30.3, 2.6.30.4+ boots intermittently for me (like someone said above, it just hangs)

I'm currently bisecting to find the bug that bfs brought out in the kernel (since it's not an actual bfs bug - it's a kernel bug that bfs happened to expose)

[predatorfreak]

Hmm, Con's made something of a comeback, eh? My good ol' buddy the SD scheduler's back.

Might have to dust off quilt and get back to this, if I can spare the time between college and my other projects.

[xmixahlx]

bfs was unusable for me but did boot up etc. (horrible lockups and i had to patch kernel/freezer.c to change #else to #endif

i switched back to using master-2.6.30

i was getting "panic early exception" messages compiling the debian way (using the official debian config as template and just adding on zen options with an initrd image).

...so i compiled everything into the kernel (master-2.6.30, zen5) and using pre-empt & cfs is running smooth.

i'm interested in the BFS work so hoping it becomes usable soon.


-michael

[martin.k]

Code: Select all

 http://ck.kolivas.org/patches/bfs/ 

It's time to dust off some "ye olde stuff"...

P.S.
Please God, just bring back Jake Moilanen with his kernel genetic stuff and I'm happy

[Ant P.]

Just updated to .31-rc8...

Holy crap... I think BFS just made my Folding@Home run 25% faster! (11 minutes per 1% -> 7½ minutes)

The downside is that it seems to break compcache... but for a speedup like this I can ignore that.

[ponciarello]

http://ck.kolivas.org/patches/bfs/bfs-faq.txt



manually applied this patch to master-2.6.30: all ok

[rahulthewall]

Fails to compile:

Code: Select all

  MODPOST vmlinux.o
WARNING: vmlinux.o(.text+0x2148db): Section mismatch in reference from the function pcibios_scan_specific_bus() to the function .devinit.text:pci_scan_bus_on_node()
The function pcibios_scan_specific_bus() references
the function __devinit pci_scan_bus_on_node().
This is often because pcibios_scan_specific_bus lacks a __devinit
annotation or the annotation of pci_scan_bus_on_node is wrong.

WARNING: vmlinux.o(__ksymtab_gpl+0x23d8): Section mismatch in reference from the variable __ksymtab_pci_legacy_init to the function .init.text:pci_legacy_init()
The symbol pci_legacy_init is exported and annotated __init
Fix this by removing the __init annotation of pci_legacy_init or drop the export.

  GEN     .version
  CHK     include/linux/compile.h
dnsdomainname: Unknown host
  UPD     include/linux/compile.h
  CC      init/version.o
  LD      init/built-in.o
  LD      .tmp_vmlinux1
kernel/built-in.o: In function `perf_counter_remove_from_context':
perf_counter.c:(.text+0x30162): undefined reference to `task_oncpu_function_call'
kernel/built-in.o: In function `perf_counter_disable':
perf_counter.c:(.text+0x30283): undefined reference to `task_oncpu_function_call'
kernel/built-in.o: In function `perf_counter_enable':
perf_counter.c:(.text+0x30334): undefined reference to `task_oncpu_function_call'
kernel/built-in.o: In function `perf_install_in_context':
perf_counter.c:(.text+0x303a1): undefined reference to `task_oncpu_function_call'
make: *** [.tmp_vmlinux1] Error 1