Official thread: "zen-sources"

[Belliash]

http://linux.derkeiler.com/Mailing-List ... /4157.html

exactly... however i used another patch for that


Also CCache is a nice feature... I used to use it some time ago...

http://linuxcompressed.sourceforge.net/

[eatnumber1]

Could You add the "USB Hid Interval Pooling" patch, as i have Logitech MX510 mice and i would like to change that from 10 to 2
I actually use 2.6.23-kamikaze5 and i had to apply that patch manually...

Thanks!

Looking at the code, it looks like that patch is already applied. Have a look at drivers/hid/usbhid/hid-core.c around line 53.

[kriko]

Could You add the "USB Hid Interval Pooling" patch, as i have Logitech MX510 mice and i would like to change that from 10 to 2
I actually use 2.6.23-kamikaze5 and i had to apply that patch manually...

Thanks!

Looking at the code, it looks like that patch is already applied. Have a look at drivers/hid/usbhid/hid-core.c around line 53.

Can you make it configurable (so you can configure polling interval from menconfig / xconfig)?

[eatnumber1]

Also CCache is a nice feature... I used to use it some time ago...

http://linuxcompressed.sourceforge.net/

It looks like that project is either dead, or it's development is too far behind the current linux kernel to be of any use. The latest patch they have is for 2.6.21... If anyone wants to port it however, i'd be more than happy to look at it.

EDIT: I took a quick stab at porting it, and some of the stuff seems non-trivial to port... especially since it was written for a locking pagecache and since then we have moved to a lockless one.

[eatnumber1]

Could You add the "USB Hid Interval Pooling" patch, as i have Logitech MX510 mice and i would like to change that from 10 to 2
I actually use 2.6.23-kamikaze5 and i had to apply that patch manually...

Thanks!

Looking at the code, it looks like that patch is already applied. Have a look at drivers/hid/usbhid/hid-core.c around line 53.

Can you make it configurable (so you can configure polling interval from menconfig / xconfig)?

I think it already is configurable via a module paramater called mousepoll.

[Spaulding]

i have some problems with kernel and gcc4.3.0 :/
1.

Code: Select all

cat: /sys/class/net/eth1/carrier Bad Argument 

and my wireless doenst work :/

[PaulBredbury]

For mousepoll, see wiki

Code: Select all

usbhid.mousepoll=2

No need to recompile the kernel for this.

[HecHacker1]

For mousepoll, see wiki

Code: Select all

usbhid.mousepoll=2

No need to recompile the kernel for this.

hmm.. I was under the impression that the usb polling feature was removed when the USB core was changed a while back to poll based on demand instead of an interval.

Does that mean it is still advantageous to configure polling time for mice and touchpads?

[PaulBredbury]

mousepoll is still used.

Code: Select all

grep mousepoll /usr/src/linux/drivers/hid/usbhid/hid-core.c

[HecHacker1]

genetics = more throughput over a longer period of time (tens of seconds); on some tests genetics is the fastest, and others the slowest... since it has to relearn the i/o pattern each time you are never guaranteed a certain latency or throughput. I would argue its best usage is for a predictable server pattern, or long file transfer.

so throughput is the only thing that matters to you?

actually, the genetic libs are very VERY customizable, right now I think all process experience the same genes, however the kernel now supports grouping of processes in many different ways so we might be able to assing a particular pid or group a set of genes or phenotypes, or give them different mutation rates, etc.

in other words, the genetic libs are implemented one of a million possible ways, the genetic libs are only limited to our imaginaion and programming skills

now if you think about it for a second, what do you actually do in real life on your computer that lasts shorter than a few seconds, every game you play im sure you are playing for a minute or longer, everytime you emerge --sync, everytime you emerge world....... emerging is by far the most cpu + disk intensive with gaming close behind and both of these usually last for minutes or longer, show me a real life task or multitasking situation where genetics doesnt kick ass!

if you can show situations were genetics sucks then maybe we can tweak things for the good.... benchmarks are highly missused, they are like laboratory experiments, how much real world validity do they have, and I NEVER sceen anyone average multiple trials together or use any sort of statistic

I am not trying to dismiss anyone's claims specifically, I just want people provide enough details so others can reproduce things..... try to make things more scientific

has anyone ever noticed how mingo and ck released "stressor" apps to test their scheduler changes so that everyone knew exactly what they were trying to do.

Well, I have a compiz/xfce desktop on a laptop, and I value low latency over throughput. According to FFSB benchmarks done by a few different people (here are just some of the google hits):
http://www.cs.arizona.edu/classes/cs630 ... atasha.pdf
http://www.kernel.org/doc/ols/2006/ols2 ... 73-180.pdf
http://www.kernel.org/doc/ols/2005/ols2 ... 35-346.pdf

The genetic algorithm takes at least 10's of seconds to stabilize for maximum throughput. I assume they are using standard genetic parameters (whatever those might be).

My own FFSB tests:
http://forums.gentoo.org/viewtopic-t-61 ... t-100.html

Show that anticipatory w/o genetic won across the board for the "mixed" workload. But I do agree that FFSB is artificial, and so the next set of tests I will use "emerge --sync" as my benchmark (since we are all so familiar with that). Also, since that test I have switched to XFS since reiser4 was causing my GUI to lag (probably because of reiser4 CPU usage). However, a "emerge --sync" on an empty portage tree is really just a bunch of writes (right?) instead of a mixed workload. Can anybody suggest a read portage workload?

If you read the documentation on Anticipatory and CFQ, anticipatory still should be the best for a single user, and CFQ is best for multiple users (or a single user with many Disk i/o's at the same time).

I value latency over throughput since compiz effects running at 65fps (my refresh rate) is 1/65 seconds or 15.38 ms. Meaning that if the scheduler can keep up at 15ms I'll never notice any lag. However, I realize that some HD seek times are just about this slow anyways.

I'd rather be able to open up "firefox, a terminal, and gedit" in a predictable low latency fashion, then have genetic sometimes open them up really fast, and other times really slow (since each time it evolves a different way). Firefox browsing is a low latency operation; the page cache has to be read before the page displays. I noticed the most lag when doing an "emerge --sync" while trying to browse the web when genetic was enabled. Firefox would sometimes get starved and go unresponsive.

I would consider gaming to be a low latency operation too, considering if you want to load a texture or sound, wouldn't you want it ASAP? AFAIK, latency in a game leads to lag. (assuming the texture or sound doesn't normally take 10s of seconds to load, and thus the Genetic algorithm will never stabilize before the loading is completed)

Of course, I didn't realize that you guys tuned the scheduler for low latency? If that is the case, I may have to try the genetic algorithms again. Maybe it was just a bad interaction with reiser4.

[HecHacker1]

mousepoll is still used.

Code: Select all

grep mousepoll /usr/src/linux/drivers/hid/usbhid/hid-core.c

So I guess its more like NO_HZ; polling based on demand up to a minimum interval. Powertop shows that I only wake the cpu when I am actually using my touchpad or keyboard (or any other usb device).

[kriko]

I would consider gaming to be a low latency operation too, considering if you want to load a texture or sound, wouldn't you want it ASAP? AFAIK, latency in a game leads to lag. (assuming the texture or sound doesn't normally take 10s of seconds to load, and thus the Genetic algorithm will never stabilize before the loading is completed)

Games are smart, loading during gameplay is done a bit in advance, so genetic will not hurt anything, at least I didn't notice anything here, though I was running another disk-consuming application (ktorrent, just some UL, enough to distract genetic) while playing, but it might help loading levels faster, since they are contained in large files.
It is true that genetics will bring you different results in same situations over time, but it is not that bad at all. For me genetics means higher data transfer with larger files (before I was reaching max 30 MB/s, average transfer rate of a large file was around 10 MB/s, which is not enough for stable DVD burning at higher speed than 6x, with genetics it will stabilize and reach 48 MB/s).

I definitely recommend it for a balanced machines between generic multimedia (games, video, music...) and desktop computing.

[Martigen]

Hey,

Love the kernel Waninkoko

Just quickly -- I moved from from zen sources 2.6.24_rc4-r1 to the git-pull 9999, and while the kernel seems even snappier during boot, I can't login to the system. GDM hangs, and switching to a console allows me to type in a username after which it just sits there. I can still switch VTs, but I'm never prompted for a password.

This is using the same .config for 2.6.24_rc4-r1, minus any automatic changes from running make oldconfig, and using stock cflags to be sure.

I saw a previous post on login problems associated with nvidia-drivers, but this isn't the problem -- again, I can't login even at the terminal once the kernel has booted. It accepts a username, and just sits there. No errors appear during boot, either.

I did actually have this problem once before with an earlier (stable-release) zen, but it disappeared with the next release, and I assumed it was a bug you had found and fixed. Alas it's back and the last week or two of git-updates to 9999 hasn't fixed it.

Any ideas?

[kriko]

@Martigen: try to boot just into console, without any graphic initialization. Have you tried with another user?

[AYBABTU]

Hey,

Love the kernel Waninkoko

Just quickly -- I moved from from zen sources 2.6.24_rc4-r1 to the git-pull 9999, and while the kernel seems even snappier during boot, I can't login to the system. GDM hangs, and switching to a console allows me to type in a username after which it just sits there. I can still switch VTs, but I'm never prompted for a password.

This is using the same .config for 2.6.24_rc4-r1, minus any automatic changes from running make oldconfig, and using stock cflags to be sure.

I saw a previous post on login problems associated with nvidia-drivers, but this isn't the problem -- again, I can't login even at the terminal once the kernel has booted. It accepts a username, and just sits there. No errors appear during boot, either.

I did actually have this problem once before with an earlier (stable-release) zen, but it disappeared with the next release, and I assumed it was a bug you had found and fixed. Alas it's back and the last week or two of git-updates to 9999 hasn't fixed it.

Any ideas?

Hi, try to recompile kernel (boot a livecd image and chroot) with zen kernel tunables set on "Gaming" setup. I think it's a dirty_ratio issue. I've got a similar issue too.
I use this custom setup:

Code: Select all

 (10) Scheduler targeted preemption latency                                                                  │ │
  │ │                                      (2000) Scheduler minimal targeted preemption granularity                                                    │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_OTHER tasks                                                      │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_BATCH                                                            │ │
  │ │                                      (50) Default round robin timeslice                                                                          │ │
  │ │                                      (66) Percentage RAM filled with mapped pages      
                                              (0) Tail large files                                                                                        │ │
  │ │                                      (1) Dirty ratio      

If i set dirty ratio on "0" value, the system just hangs on x start, and i can't compile anything.

[Martigen]

Hi, try to recompile kernel (boot a livecd image and chroot) with zen kernel tunables set on "Gaming" setup. I think it's a dirty_ratio issue. I've got a similar issue too.
I use this custom setup:

Code: Select all

 (10) Scheduler targeted preemption latency                                                                  │ │
  │ │                                      (2000) Scheduler minimal targeted preemption granularity                                                    │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_OTHER tasks                                                      │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_BATCH                                                            │ │
  │ │                                      (50) Default round robin timeslice                                                                          │ │
  │ │                                      (66) Percentage RAM filled with mapped pages      
                                              (0) Tail large files                                                                                        │ │
  │ │                                      (1) Dirty ratio      

If i set dirty ratio on "0" value, the system just hangs on x start, and i can't compile anything.

Yep, that worked! Thanks

Wonder what the issue is with a 0 percent dirty ratio that causes it to segfault so bad.

Do we have documentation on the properties of the tunables?

[Waninkoko]

Hi, try to recompile kernel (boot a livecd image and chroot) with zen kernel tunables set on "Gaming" setup. I think it's a dirty_ratio issue. I've got a similar issue too.
I use this custom setup:

Code: Select all

 (10) Scheduler targeted preemption latency                                                                  │ │
  │ │                                      (2000) Scheduler minimal targeted preemption granularity                                                    │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_OTHER tasks                                                      │ │
  │ │                                      (1) Scheduler wakeup granularity for SCHED_BATCH                                                            │ │
  │ │                                      (50) Default round robin timeslice                                                                          │ │
  │ │                                      (66) Percentage RAM filled with mapped pages      
                                              (0) Tail large files                                                                                        │ │
  │ │                                      (1) Dirty ratio      

If i set dirty ratio on "0" value, the system just hangs on x start, and i can't compile anything.

Yep, that worked! Thanks

Wonder what the issue is with a 0 percent dirty ratio that causes it to segfault so bad.

Do we have documentation on the properties of the tunables?

I think I know where's the problem. I'll try to fix it.

[AYBABTU]

I think I know where's the problem. I'll try to fix it.

Good, many thanks!

Could you please inform us when (take all the time you need ) this problem is solved? So we can set "Low Latency Desktop" without any issue

[RobertDavid]

Does someone have errors on AMD64 X2??

Since rc6-zen0 I heve some "SMP dividing error" randomly ocures while using system (first on boot, second mainly running some bigger program like X od compilation), looks like freezing programs that maybe try to use both processors ("mc" works ok w/o problem),,, rc5-zen0 is fine and have same config. I will paste the exact error later (maybe you know what is it).

I still dont have my kernel config fine tuned for my system whyle I have that only since last week.

My config:

http://robertek.brevnov.net/files/linux/config

[Dottout]

I think I know where's the problem. I'll try to fix it.

Good, many thanks!

Could you please inform us when (take all the time you need ) this problem is solved? So we can set "Low Latency Desktop" without any issue

you can look at http://repo.or.cz/w/linux-2.6/zen-sources.git to see summary

[ilikenwf]

Well...with C1E disabled so that dynticks and hr timers work, alongside the powersaving stuff, I am really surprised at how long I'm lasting on battery...wow!

Error on current source:

Code: Select all

make[6]: *** No rule to make target `/usr/src/zen-sources/drivers/net/wireless/madwifi/ath_hal/../hal/public/x86_64-elf.hal.o.uu', needed by `drivers/net/wireless/madwifi/ath_hal/x86_64-elf.hal.o'.  Stop.

Looks like the madwifi drivers aren't even in the /net/wireless folder.

[eatnumber1]

Well...with C1E disabled so that dynticks and hr timers work, alongside the powersaving stuff, I am really surprised at how long I'm lasting on battery...wow!

Error on current source:

Code: Select all

make[6]: *** No rule to make target `/usr/src/zen-sources/drivers/net/wireless/madwifi/ath_hal/../hal/public/x86_64-elf.hal.o.uu', needed by `drivers/net/wireless/madwifi/ath_hal/x86_64-elf.hal.o'.  Stop.

Looks like the madwifi drivers aren't even in the /net/wireless folder.

There may be a problem with your tree if there is no madwifi directory in drivers/net/wireless. Regardless, try posting your .config file and i'll take a look.

[ilikenwf]

http://www.mattparnell.com/zen/2.6.24-r ... md64config

[eatnumber1]

http://www.mattparnell.com/zen/2.6.24-r ... md64config

I'm having some trouble testing your kernel... Check your tree however, since the tree definitley has the driver... See here

[ilikenwf]

http://www.mattparnell.com/zen/2.6.24-r ... md64config

I'm having some trouble testing your kernel... Check your tree however, since the tree definitley has the driver... See here

Will try tomorrow...am on the low bandwidth connection at home now...