With feedback, Ingo Molnar has continued to improve his voluntary kernel preemption patch. Testing the patch has revealed a number of areas in the 2.6 Linux kernel that were causing high latency. Fixes have been created and merged as these areas have been located. Read more at KernelTrap.
Reducing Linux Kernel’s Latency
About The Author
Eugenia Loli
Ex-programmer, ex-editor in chief at OSNews.com, now a visual artist/filmmaker.
Follow me on Twitter @EugeniaLoli
First post yaay
No seriously guys, I am using 2.6.8.1 & the sound still breaks whenever updatedb is started
I mean 2.6 kernels where supposed to be ‘desktop friendly’ & blah blah..
Hope this new patch helps.
Is it true the 2.6.8 family is the crappiest 2.6 kernel?
Also anyone knows when will reiser4 make it in the official kernel?
2.6.8.1 is dodgy for me as k3b won’t work, the sound bearks up under load and the system crashes a lot more than usual
Roll on 2.6.9!
I have been amazed at how nice 2.6 has been on Fedora 2 and SuSE 9.1. Not sure which version of 2.6, but great work.
“No seriously guys, I am using 2.6.8.1 & the sound still breaks whenever updatedb is started
”
It seems to me that it’s updatedb that should be fixed
I would prefer that it ran 10 min instead of 1, but took 10% of the resources it takes now… I really hate updatedb.
Would someone clear up the confusion that I have on whether or not the 2.6 kernel is a development or stable branch. I was under the impression that it was the stable branch at the beginning but now since I don’t see a 2.7 kernel in the works and how the kernel hackers are added/modify features in the kernel that are fairly significant in the 2.6 series … I start to think that 2.6.x is now developmental and not a stable branch.
Can someone please explain the kernel numbering systems.
Thx
Its interesting how its working.
usally its 2.even stable
2.odd development
but they made a decision to put progress into 2.6 :/
guess they are leaving it to the distros
” I am not that much expert in testing latencies, but I can say safely that Linux kernel needs a very long time to be as fast as windows xp. ”
I am extremely happy with the 2.6 kernel compared to xp. I no longer dual boot. This is of course in a i586 kde setup.
2.2 is stable
2.3 not stable / testing
2.4 is stable
2.5 not stable / testing
2.6 is stable (or supposed to be).
I remember reading an email stating that the work on the 2.7 kernel (testing) was put on hold. And all the developers were going to be merging in changes / patchs into the stable release. IMHO, 2.8 should not have all these drastic changes put forth. It seems like it is causing some instablilities. Most major distro have not put 2.6 into the stable realm yet. For example:
1) Fedora is the testing platform and it is the 2.6 kernel
2) Slackware is using 2.4
3) Debian (sid / testing) is using 2.6, I believe sarge is 2.4
4) Mandrake is using 2.6 now but they have always put new features in before the other distros.
5) Suse is running 2.6 now. I think they have patched their own kernel.
Personal opinion here, the developers were put off from doing 2.7 until various issues get hashed out of 2.6. Once 2.6 gets the kinkgs ironed out then the 2.7 branch will be created (imho).
“Linux Kernel is very unstable while the OS higher layers (Services, GUI and others )are much more stable.”
i’ve got 2.4.26 egg & don’t know what are You talking about – system craches? – never ! i find linux kernel extremely stable (maybe there are some unstable versions , perhaps 2.6.8 , but i am sure that rock solid one’s are also available)
i only experienced 2 nautilus crashes in the past 3 months , nothing that would affect kernel directly ; i think , if anything is unstable that would be those higher OS layer hraq mentioned (like nautilus , xine)
“I think that linux kernel development is running very slowly, and all that we see right now is evolutionary at best rather than evolutionary.”
i thought it is under heaviest development in he open source
2.6.8.x has been unusable for me. It does not connect
to certain websites on my setup. But 2.6.9-rc1 and 2.6.7
are excellent.
Well,
Whats your definition of stable? 🙂
Historically I can tell you that it sorta goes like this:
2.4.0-9 Release (Suppose to be stable)
2.4.10 Production Releases (Truly are been in the wild for about a year…)
2.6 is stable, what you are seeing is discussions not about stability but performance, which I think is a separate issue.
This is good. It means we are moving beyond the bugs of the initial release and are now concentrating on improving the engineering of the code with respect to performance.
Not everyone is fully supporting 2.6 yet. I suspect in another year 2.6 will come into its own and the learning curve for user space code will have been completed.
You should see more stability in stuff like sound support, desktop support in X, etc.
As for the comparison with Windows, Windows will eventually fall by the wayside.
We have the source, and we can improve our kernel, even get feedback from users.
Try doing that with Windows.
We WILL prevail.
Besides…2.7 is just around the corner.
Watch it Billy Boy, My Plush Penguin Stuff Toy in my server room is gunnin for ya.
🙂
-gc
2.6 is considered the stable branch while the 2.6-mm patchset is considered unstable (instead of 2.7)
“Is it true the 2.6.8 family is the crappiest 2.6 kernel?”
Not on this desktop.
“Also anyone knows when will reiser4 make it in the official kernel?”
I don’t know about official, but there are reiser4 patches for Gentoo which work like a charm.
I had the same problem you had with broken sound. It would work, but whenever I got a message in gaim it would just crack and either be broken or go into a repetitive loop. I had to lower the sound through the alsamixer, I believe I had to lower the master and pcm controls down to about 70%. After that it worked fine, no more broken up sounds and no more repetitive looping sounds.
I have been using the 2.6.x kernel on Gentoo for quite a while (as long as it has been in gentoo-dev-sources … since I think shortly after the 2.6.x release).
I have this kernel specifically on a Multimedia machine, that I use for editing (and rendering) digital video from my Cannon MiniDV Camcorder and for recording TV digitally via MythTV.
I haven’t seen any performance issues on this machine and 2.6.x (currently, 2.6.8-gentoo-r4) … but I don’t run updatedb while I am doing video work.
> sound still breaks whenever updatedb is started
Don’t you think that you just need to tweak IO scheduler ?
, or at least but not best reduce updatedb nice level.
> 2.6.8.1 is dodgy for me as k3b won’t work
run the command:
chmod +s `which cdrecord`
and it will probably work again. In 2.6.9 this new “feature”(too long to explain here) will be the same.
> whether or not the 2.6 kernel is a development or stable branch.
It is stable branch, but is accepting not only bugfixes but also lots of changes. Linus says kernel stablized in spite of lots of changes.
As for the posting By hraq (IP: 69.111.190.—)
It seems some microsoft fan. ALL his statings are FALSE.
Especially “linux kernel development is running very slowly” – this is ridiculous, he is either lying, or disconnected from reality (100% INcompetent)
“It seems to me that it’s updatedb that should be fixed
I would prefer that it ran 10 min instead of 1, but took 10% of the resources it takes now… I really hate updatedb.”
On my debian/sid /usr/bin/updatedb is just a shellscript which is running “find” over the complete filesystem. Inspired by your comment I just added
SEARCHPATHS=”$SEARCHPATHS -exec sleep 0.004 ‘;'” (line 120) which approximately makes updatedb run 10 minutes longer because it waits 0.004 seconds for each of the 147580 files which are currently installed (cat /var/lib/dpkg/info | wc -l), note that it might be “sleep 0,004” for you depending on your locale, it would probable be safest to do “export LANG=C” to make the dot work.
CPU usage is at 20% now as opposed to 100% and sound is no longer skipping
I am running the 2.6.8.1-7 kernel in Yoper Linux V2, and it is very stable infact nothing has ever locked up or crashed. This is the third 2.6 kernel I have run under this distro, and I have had no problems at all. The audio is kind of strange though, what ever app is using it, the other ones have to wait to get access to the sound server, they all store the sounds, then when I shut Xmms down I hear all these blasted at you at the same time. Why can’t all these sounds be mixed and sent to the sound server at the same time.In any case Yoper Rocks! Oh and is very fast!
Another alternative is using nice. running “nice -n 10 updatedb” will give you a cpu usage of nearly 100% except when you need it for a nice 0 process.
The audio is kind of strange though, what ever app is using it, the other ones have to wait to get access to the sound server, they all store the sounds, then when I shut Xmms down I hear all these blasted at you at the same time. Why can’t all these sounds be mixed and sent to the sound server at the same time.
You’re probably running things under ALSA with a sound board that has only one channel. Thus, you have to enable software mixing. See here how, it’s just a small config change:
http://www.fedoraforum.org/forum/showthread.php?t=18522
And yes, I do agree that it’s silly that you have to do that manually, I saw the issue discussed on the fedora and redhat mailing lists and hopefully it will be automatically set correctly in the next releases.
This is about Ingo Molnar and his voluntary kernel preemption patch. I’m running gentoo with the vanilla kernel patched with volutary preempt. Audio latencies are very low and stable. I’m not talking about mixer settings, I’m not talking about boot times, I’m not talking about disk IO (well not directly). I’m talking about low latency in terms of sustained short round-trip audio throughput (this isn’t the only way to measure kernel latency but it’s a heck of a lot better than ” It does not connect to certain websites on my setup”).
The real result is that Ingo Molnar and a few other devoted latency detectives are scouring the kernel for rough spots and smoothing them out.
The outcome? Well I can get audio latencies (using jackd) of around 3ms! Unless you work with professional audio applications, your jaw may not be on the floor… but for those of you who know windows/mac audio apps well, you know around 3ms of latency is pretty amazing.
No, it’s already started with a nicevalue of 19 (export NICE=19 in /etc/updatedb.conf) and music playback still stutters. Also I think I/O, not CPU is the bottleneck here because kernelcompiling does not show such sideffects at all. XMMS should buffer more I guess
It’s worth noting that this only really benefits pro-audio, and then only when low latency is required.
If you don’t need low latencies, then just use larger buffers and hardware monitoring. I get rock solid 24 track record/playback with an unpatched 2.6.8.1 with 1024 frames per period. (About 23ms one way)
If your mp3s are skipping, set your buffer size larger in the sound server, or try another mp3 player. It is of little importance whether you hear your music 500ms or 10ms after pressing play. Large Latencys will not cause film soundtracks to lose sync as the video player compensates.
There are problems with audio in Linux, but as with many things Linux, they are further up the stack than the kernel, and would perhaps be better considered a configuration and desktop interface challenge. (Which of the plethora of sound servers to use and which apps support which server, ALSA config files and silly default settings.)
@ hraq
Low latency under Windows is achived by forcing soft real time scheduling and obscenely high priority. Linux at the moment does essentially the same thing when you launch jack with -R.
In theory though, the ‘correct’ way is to improve the schedular and minimise time spent in the kernel+spin locks so that this is not required.
This avoids the situation where an audio app can crash and totally lock the computer, as no other task can have priority over it. It also avoids having to let applications raise their own priority (normal user processes can only lower their priority under Linux, which is safer).
The Windows way is ‘good enough’, but for some reason Linux kernel hackers have never been happy with compromises when there is an engineering solution. That is the focus of this patch.
How did is mesure the 3ms ?
2.6.8.x has been unusable for me. It does not connect
to certain websites on my setup.
A kernel does not connect to bloody web sites.
The kernel loads the module which talks eth0 card which
talks to the cable modem provider (Unfortunatly NTL) The
kernel is responsible for network communation on some level.
Although my comment should realy by modded down in hinesite
it still does not solve the problem with reduced throghput
on version 2.6.8.1 when it is fine on 2.6.7. There is no
strange msgs in the system logs that help me discover the
problem. Other users have already issued a report on bug
numbers 3247 and 3212. Is the any point of surfing
@ 42bytes per second ?
I am a fan of Linux kernel that the rest of my software sits
on and have moved away from windows. But it does not mean
that its made inerrantly. Hopefully I will be able to contribute code in the future. But now I can only report
bugs I notice. The best thing about Linux is that the
problem will be solved probiliy in the next version.
so far 2.6.8 is the “worst” release for normal users:
for .8, NFS was broken and .1 was released.
still, cd burning as non root and sound problems exist…
hopefully .9 is released with better care
> A kernel does not connect to bloody web sites.
Wtf? Where else do you think the TCP stack, and especially connect(), lives?
The original poster’s problem is likely related to ECN (Explicit Congestion Notification) being enabled in his new kernel. There are still a lot of broken routers around that don’t support this. It can easily be disabled by echo 0 > /proc/sys/net/ipv4/tcp_ecn
Gnome theme changes into heavy imaged themes also cracks the sound, and it doesn’t with a 2.4 kernel.
That and updatedb are the only ways I’ve been able to crack it, so it’s not much of a disturbance. 16 hours in a day, and I’ll probably hear two cracks in it if I were to listen all day, and they’re from the updatedb cron job at midnight.
kind funny to think about it, windows does sound mixing on the kernel by ages, but linux is even now with this sound dilema, sound mixing when no hardware support exists, ioscheduling and sound latency etc. lets see if alsa has came to stay
about the net prob, i got on same extent some problems related to, trying to connect to some sites, at first (and random ones) none seems to connect, after awhile i could, never thought could be kernel related. Always thought it was the isp, dns/router, my adsl line? who knows, but earing from others and after changing kernel, only 6.8 was the faulty one to point
i have some probs too with the latest kernel and nvidia drivers, x seems to refuse to cooperate and the enigmatic rcharbsod (random chars on black screen of death) appears
and another strange insue i had in this past few weeks was the lack of suport for codepages on the vfat fs on 6.7, its fixed on 2.6.8, the funny thing is why that happened… vfat fs is stalled for agesss!!
i guess 2.6 isnt stable yet for mass usage. my predictions of stabillity (as alwasy) would be when we only approach 2.6.20. Until that times came by… well keep a 4.27 around and keep playing with latest andrew patches
What is your justification for that statement? In my experience, Linux behaves much better under heavy VM and I/O loads than Windows. It’s exhibits it’s best performance under light desktop workloads.
“I don’t know about official, but there are reiser4 patches for Gentoo which work like a charm. ”
Those patches are for the kernel, not gentoo. Please stop pointlessly pushing gentoo.
“Those patches are for the kernel, not gentoo. Please stop pointlessly pushing gentoo.”
Uh, the patch applies to gentoo-dev-sources, which IS the Gentoo-patched Linux kernel, presently at version 2.6.8-gentoo-r3. The reiser4 patch applies cleanly and works
If the sound breaks while running updatedb in 2.6 then there is something wrong with your configuration, sound drivers or you are using a sound server like ‘arts’. Im running a 2.6 kernel with ALSA EMU10k1 driver and I can start compiling Xorg, start updatedb 5 times and watch a DVD in xine all at the same time with no sound skipping at all, of course im not using a sound server like arts.
While watching a dvd in windows i cant even install and application without it skipping really bad.
I’m using 2.6.8-r3 of gentoo-dev-sources, and I’ve had quite a lot of problems with sound skipping, especially when there’s heavy disk usage.
Not to mention how sound totally stalls out for a few seconds when I switch between framebuffered terminals at different resolutions.
Rayiner Hashem: It’s the worst behavior I’ve seen under 2.6, by far.