Cliff White recently posted some re-AIM multiuser benchmark results comparing the stable 2.4.23-pre5 kernel against the 2.6.0-test5 and 2.6.0-test5-mm4 development kernels. In his conclusion he makes reference to earlier scheduler tests posted by Mark Wong saying, “Short summary: we mostly rock.“
Linux: Benchmarking 2.4 vs. 2.6-test
About The Author
Eugenia Loli
Ex-programmer, ex-editor in chief at OSNews.com, now a visual artist/filmmaker.
Follow me on Twitter @EugeniaLoli
I read at the udev pages that it aint ready yet for 2.6 prime time.
devfs doesn’t work with 2.6.
-magg
… is the new scheduler suppose to address ‘responsiveness’ or better scalability (or both). I have seen high benchmark numbers for various kernel releases but if playing an mp3 still stutters than what good is it?
thx
u know what i wanna know… i wanna see some comparative benchmarks of these 2.6-test kernels versus freebsd 5…
ive heard a lot about the huge improvements from 2.4 to 2.6… and everyone seems to be comparing just those two kernels… but my question is how does it stand up to the bsds… heck, y limit yourself to freesd; test net and open too while u’re at it…
eugenia, r u listening <or reading, since u r not actually listening….>
do u wanna run some benchmarks, eugenia? i dont have a computer to spare to do them myself
devfs doesn’t work with 2.6.
That’s a feature not a bug.
🙂
@anonymous: You have to give the sound-playing process a higher priority if you don’t want your MP3s to skip. MP3s skipping hasn’t been a problem for me ever since the preempt patches came out, and I used 2.4 on a PII-300 at the time. And don’t get the idea that giving the sound-server a higher-priority is somehow cheating. That’s exactly what the priority mechanism is there for — to tell the kernel which processes are more important than others. WinXP does the same thing — it gives a +8 priority boost to any process using an audio device, while BeOS has special scheduling priorities for audio processes. If you use KDE, aRts should get started with real-time priority anyway, so you should not get skipping problems. I use 2.6 on my laptop, and I don’t get audio skipping even with multiple compiles going on in the background.
Also, the new scheduler is supposed to address a lot of things. On the desktop side, it is meant to increase responsiveness, through the integration of the preemptive patches. It is a huge success on this front: I can run multiple compiles in the background and not even notice an impact on GUI responsiveness. The new I/O scheduler helps a lot here too. It solves a problem that really shows up in Windows a lot — an interactive task that needs to do disk I/O (like launching a program or opening a file) often has to wait a long time for a background task’s I/O to complete. This makes a big difference in how fast the system feels under load. On the server side, the new scheduler is supposed to enable the usage of huge amounts of threads (hundreds of thousands) because the CPU time used by the scheduling algorithm is not proportional to the number of threads. Also, its supposed to allow the kernel to make better use of larger numbers of processors.
Um, devfs appears to be working fine over here. I’ve got kernel 2.6, and the /dev/input/mouse0 link I use for my USB mouse is still there…
DevFS do work now, but it’s deprecated. It’ll be removed in 2.7.
Rayiner, you’re using Gentoo, right? I heard devfs is broken in -test5. It was okay with -test4 though.
I’m not a desktop speed junkie, but I did notice a definite improvement in responsiveness. I am grateful for two things, though:
1) The CD burning code. It’s so nice to just use ‘cdrecord dev=/dev/hdc’, especially since I can now use hdparm to tweak the drive options. (if only CDRDAO would support it!)
2). Qtconfig is a great program. It’s pretty, and it displays the elements of configuration all in one screen. Very helpful.
All in all, I’ve noticed several improvements in the general polish of the kernel administration. Minor things, mostly–like the concise gcc commands it prints to stdout when compiling. Also, it seems to be much speedier–I compiled it on a PPro 180 in only a few hours.
Since RedHat is not going to release its free version in the same name(redhat), why don’t they release a preversion with 2.6 kernel in it that has a auto bug reporter
Microsoft does that even in their mainframe OS’s+internet browsers, and they are not shame of doing that, we pay them, we help them find bugs in their OS, we help them develop better OS’s and browsers..this way they prepair patches for their buggy OS and sell the bug fixes we found back to us and we pay them again(they usually call those bug fixes SECOND EDITION or ME:)
why don’t we do the same for linux.
The new I/O scheduler helps a lot here too. It solves a problem that really shows up in Windows a lot — an interactive task that needs to do disk I/O (like launching a program or opening a file) often has to wait a long time for a background task’s I/O to complete.
Yesterday I was formatting a partition in Windows. At the same time I tried to surf the Internet. There were regular hickups that lasted for five to ten seconds. So it was pretty much useless.
However, I noticed that it depends on the program as well. At the same time I downloaded a file using WS_FTP and that program was very responsive. So was Outlook Express. IE and Explorer on the other hand were next to useless because they worked for only a few seconds before the next delay came.
I hope things like that won’t happen with the (new) Linux kernel.
I use the vanilla kernel version 2.4.22. There are just two things you need to do improve responsiveness of your gui while under heavy load such as compiling. This trick is particularly pertinent for source based distro users, for obvious reasons.
Trick 1
Decreasing the niceness of X. I have X set to -20 on my linux boxes. I can’t overemphasize the speed boost in responsiveness to my GUIs. It is short of amazing. Use it with caution though, I’ve heard it’s not a good solution.
Trick 2
This is particularly useful for source based distros with an organized package management system. In Gentoo’s portage, for example, there’s a option to set the niceness of portage (package manager). I promptly adjusted mine to 20. This undoubdetly may increase compilation times. But compiling CPU intensive packages have almost zero effect on UI responsiveness. In fact, from my observation, they have none.
On the other hand, when I need all the CPU power I can get for compilation, I switch to console mode, shutdown X/KDE/whatever, add -j4 to the make option, and I set portage ( the package manager) niceness to -20. Packages of course compile faster. I do this when I don’t need to use the computer and I’m compiling large packages.
So if your source based package manager supports adjusting its niceness, do not hesitate to make use of it especially if you are one like me who has 15 apps open on 15 different virtual desktops while compiling the latest version of Openoffice and listening to punk rock from juK without any skipping. If it doesn’t file a bug report to your devs to include it. It’s a life saver.
Once again, this is for the vanilla kernel which doesn’t have any of those desktop performance patches, I’ve found these tricks impressive. I await the vanilla version 2.6.
You wrote,
>I can run multiple compiles in the background and not even notice an impact on GUI responsiveness.
I realise you may have a faster CPU than me but what about, the time it takes for a program to launch.
I wonder how 2.6 will help with my system, things that may be negligible on a P4/Athlon are a nightmare on my 450Mhz K6-2. I guess those problems will need to be solved at the application level.
Ahem.
That was a joke, I haven’t had particularly good experience with devfs.
I’m glad to hear it’s being removed from 2.7.
How do you “Decreasing the niceness of X” anonymous? Please don’t be intentionally obscure.
I also have simliar problems although my computer is a little bit faster. I have a Pentium 3-m (1ghz) with 512mb RAM, yet it seems like apps take forever to launch. Mozilla and OpenOffice would take 3-5 longer to start (stopwatch measurement, in and out of cache). I will try a distrobiution with Linux 2.6 whenever they are ready, and see if this improves at all.
“Decreasing the niceness of X”
1. You should open a terminal and type top.
2. Look for the pid number of X.
3. Press q to quit the top program.
4. type su – (for root access, enter password)
5. type renice -20 pid, eg. renice -20 2345
If you are using Redhat 9, it is recommended you don’t do this to the X server. I did it and noticed no difference, on Mandrake and Gentoo you notice some difference.
“1. You should open a terminal and type top.
2. Look for the pid number of X.
3. Press q to quit the top program.
4. type su – (for root access, enter password)
5. type renice -20 pid, eg. renice -20 2345 ”
or (might be some syntax errors in it, just typing this from memory)
1. Type su – enter password
2. Type renice -20 -p $(pidof X) (something like that
)
3. If all goes well you put this in some script after x starts so x gets reniced after each reboot
Hi!
Guys, there r options
Eg i really suggest u should try MozillaFirebird on a slow computer, not mozilla. I have a fast computer, but i prefer it to mozilla. It starts 2-3 times faster than mozilla, and has the best rendering time i’ve ever seen on any platforms/browsers (o course gui browsers:). It recognizes and uses all the mozilla plugins (java,flash), and overall better than mozilla. Give it a try!
OOo 1.1rcX is better again than 1.0, they made noticable improvement in speed. If that’s still too slow, u should try abiword or other smaller packages for word processing. And for the desktop: xfce (version 4). It’s small, fast, nice, elegant, and knows everything u need from a desktop.
bye, hirisov
You should try Konqui from KDE3.2. I bet the rendering times are even better than those of Mozilla-Firebird;-)
There should be directory called xdm in in your X11 directory in etc. (i.e /etc/X11/xdm) The location of the xdm directory varies from distro to distro. In that directory, look for Xservers and add the following line to it.
:0 local /usr/bin/nice -n -20 /usr/X11R6/bin/X -dpi 100 -gamma 1 :0
Adjust all other values as you see fit. The important phrase their is /usr/bin/nice -n -20. That changes the niceness of X. Now, I’m assuming that you boot directly in X/KDE/Gnome/*windowmanager*, via a login manager like kdm, gdm, or xdm.
If you do not boot directly in X/KDE/Gnome/*windowmanagers* via kdm/gdm/xdm then use the suggestions posted by other responders above.
I hope I’m less obscure this time around.
MozillaFirebird over Mozilla on Windows, I say yes.
On Linux, not much difference, it is the XUL interface. Galeon and Epiphany are much better, I wish there was a KDE browser with gecko embedded, but obviously Konqueror is a better option there.
Although i am not Mr. Anonymous, I thought I would give you an answer.
Under Linux in the dsktop(KDE/GNOME/Fluxbox/Waimea/E17/XFCE4/FVWM/XPDE) of your choice, open a terminal(konsole,gnome-terminal-xterm,aterm,eterm,xfterm etc.)and type the following:
->] ps -Al | grep X
This will return a line which looks like the following:
4 S 0 2188 2183 1 75 0 – 84000 schedu ? 00:17:04 X
The fourth column refers to the PID(process ID). this number of course will be different than listed in this example,if you then type the following:
renice -20 -p 2188
X will be give a priority of -20, which counter-intuitively is the highest priority
likewise, if you type in the following:
renice 20 -p 2188
X, or any other app referenced by its associate PID will be given the lowest priority…which means the process will only be serviced once all other process have been serviced.
This is known as changing the “niceness” of a program. Additionaly on can simply start a program with higher/lower priority with the following command:
nice -n 20 make
this will start ‘make’ with the lowest priority(usefull for example when compile the the kernel)…
nice has the following syntax:
nice [-n adjustment] [-adjustment] –adjustment=adjustment] [command [arg…]]
whereas renice has the following syntax:
renice priority [[-p] pid …] [[-g] pgrp …] [[-u] user …]
where -20 is the highest and 20 is the lowest priority value….
I have two pc’s here running gentoo, when I want to install the latest and greatest, for example gnome-2.4, I simply use distcc and distribute the compilation between the two machines, allowing me to renice the installation process(in this ’emerge’) to the lowest priority, which means I can then do many other CPU intensive things, like playing a game, watching a movie, creating complex documents etc. The compilation, due to distcc, still proceeds quite rapidly, and my PC is still fully usable. With this approach I have virtually zero downtime due to compilation time-many lament the slowness of compilings programs from source, but for me, my machine never sleeps, is always doing something, and it remains responsive. The only “downtime” due to compilation is when doing a fresh install of gentoo on a new machine-and only until X and gnome have been compiled(maximum 24hrs.)Personally I have never had this problem of xmms skipping while playing songs, unless the soundcard I was using was very,very badly supported-my main PC is 1.3Ghz P4 w/768 mb ram with a SB Live! soundcard, a relatively new machine, but still nothing to compare to 3.0+ Ghz newer processors….
regarding the 2.6 kernel
I am in the process of making the transition to the 2.6 kernel as I am writing this. I am using 2.6.0-test5-mm3(under gentoo mm-sources). I made a series of small batch files (bash scripts)to help switching back and forth between 2.4.20 and 2.6. One batch files (re)set a symbolic links pointing to /etc/modules.autoload, /etc/conf.d/local.start, /usr/src/linux, and /boot/System.map. ie. I have local.start.2420 and local.start.26 and local.start is a symbolic link to one of the two, the same for the other files listed here. Then I have another batch file which (re)emerges criticial packages which provide kernel modules (alsa-driver, nvidia-kernel, iptables). Lastly I wrote a little script which touches all *.o and *.ko files (ie. modules) under /lib/modules/’uname-r’which prevents portage from removing the old modules when emerge the same package for the other kernel….
As part of this upgrade process I swithched to gcc3.3.1(propolice) and glibc 2.3.2 with NPTL support.At this point I have only recompile a couple of programs(X,gtk+2.4) so I have yet to see the threading improvements-(I am still waiting for a sun-j2sdk(1.4.2.01)which hopefully will fully utilize the pthreads implementation in gcc/glibc and the 2.6 kernel). At this point I see no real speed differences between the two kernels- a gentoo-sources 2.4.20 kernel(r7) with a variety of patches, but whithout preemptive multitasking and without the new (0)1 scheduler…and the 2.6.0-test5-mm3. As of this writing there is still no fritz capi module available for the 2.6.0 kernel-which means I can’t use my ISDN card as a ax/answering machine. But otherwise all of my hardware is correctly supported- There are still problems with directfb and vesa: I use ‘qingy’ as a substitute for ivgetty-instead of a normal text console I am present with 1280×1024 graphical login(without gmd/kdm/xdm). Under 2.4.20 I can start my desktop from ‘qingy’ but under 2.6 the machine goes into la la land(and doesn’t come back). I also must use an iptables ebuild from breakmygentoo.net in order to setup ip masquerading and ip transparency for my local net. But my system is now running for over 72 hours with the new kernel and it is remarkably stable. Occassionaly I get ‘mouse disorientation’ when playing games and must quickly switch virtual consoles and return to re-initialize the mouse-but I had this problem with 2.4.20 too(appears to be a problem with the vesa fb drivers in conjunction with xfree and the nvidia-kernel)…I have compiled the kernel with the following cflags:
-march=pentium4 -O3 -pipe -ffast-math -mfpmath=sse,387 -msse -msse2 -mmmx -fprefetch-loop-arrays -maccumulate-outgoing-args -falign-functions=4
which sofar has yieled excellent results….
hdparm -T -t /dev/hda
929.21
35.53
hdparm -T -t /dev/hdb
1004.15
24.75
The number seem oddly high-under 2.4.20 I never got more than 400MB/s….
glxgears remained constant between both kernels
..and even BeOS 4.5.x beats Linux 2.2, 2.4 and apparently, 2.6 for interactivity and responsiveness, even under the highest multimedia (not just mp3) load.
Poor Linux. The more it changes, the more it stays the same.
How much of the slowness is just plain ass gui bloat? Gnome and KDE are both huge. On my toshiba laptop(1.8GHZ Celeron/512MB menory)with NetBSD 1.6.1 which i changed from Redhat 9/Gnome, i use blackbox and everything smokes compaired to the redhat install. Everyones preference though I like fast and simple versus slow and bloated. The question is how much of it is OS layer versus slow GUI’s.
2.6 probably will not be the speed demon until 2.6.10. Speed is good but being stable has to be a higher priority. If you want the speed wait for WOLK (wolk.sf.net) to come out with the 2.6 kernel!
Hey thanks mtdman, I actually got a better understanding of stuff that I didn’t know that well.
X is responsible for rendering/drawing the windows on your desktop, not blackbox, fluxbox, or whatever minimalist window manager you claim to be the fastest and greatest.
The term KDE is slow is essentially meaningless. It is a lousy lie. Theoretically speaking, kwin (KDE’s windows manager) is just about as fast as fluxbox or whatever minimalist window manager you use. Why? well because neither of them are responsible from drawing the figures you see on your desktop.
On the same machine, items wouldn’t draw any faster or slower if you used KDE or Fluxbox. You systems responsiveness also remains the same. The function of a window manager are very basic. Maximize, minimize, resize, close, move, provide handle, encapsulate window. Window managers are not responsible for speed performance issues many people tend to attribute to it.
So use whatever window manager you like. If you need responsiveness on your desktop look for good kernel patches or renice X. Going minimalist doesn’t make your desktop responsive, it’s another dumb myth.
Actually, it does: it’s called memory usage. If you use Blackbox instead of KDE, less memory will be used, which means that your system doesn’t have to swap stuff all the time, making it more responsive.
Of course this doesn’t matter for modern systems that have a lot of memory.
Actually, it does: it’s called memory usage. If you use Blackbox instead of KDE, less memory will be used,
That depends on several factors outside the window manager it self. For example, the size of the apps you are running. How apps you have open. Your system specifications and so on. The window manager you use has a neglible effect on memory usage. The types of apps you run do.
which means that your system doesn’t have to swap stuff all the time, making it more responsive.
Linux usually swaps stuff that aren’t needed but that Linux anticipates will be used. Application data however is usually first stored in memory, less important data is stored in disk buffers also in physical memory, then even less important data is stored in disk cache still in physical memory. Data that will probably never be used by Linux anticipates might be used are then stored in swap.
In fact, you could have data in swap while you have some free physical memory. I’ll give an illustration. I have about 12 applications open. Although my sensors are telling me I only have 16MB of my 256MB free, it is not entirely correct. The memory actually containing application data is only 100MB filled. So technically speaking, I actually have 156MB of free RAM that the system can use.
However, Linux anticipates that I might need some data. So to speed up the system, it adds this less important data in memory called the disk cache and disk buffers, which add up to about 140MB of memory usage, their by leaving me with 16MB of free memory. Even though I have 16MB of free memory, I am still using about 7% of swap.
I’m not going to delve into shared memory with KDE and heavy weight managers make use of will could also distort memory footprints. But I hope you get the gist of the whole memory management thingy.
Of course this doesn’t matter for modern systems that have a lot of memory.
Except you use anything below 500Mhz of CPU power and 128MB of RAM, I see no reason why anyone will be using minimalist window managers, except again of course for personal reasons and taste. Every other reason is probably a myth spewed all over the internet.
If you have anything resembling a mordern machine, you are better of using a full featured desktop environment, rather than wasting precious CPU cycles and system resources.
Wow! alan t said:
Microsoft does that even in their mainframe OS’s+internet browsers, and they are not shame of doing that, we pay them, we help them find bugs in their OS, we help them develop better OS’s and browsers..this way they prepair patches for their buggy OS and sell the bug fixes we found back to us and we pay them again(they usually call those bug fixes SECOND EDITION or ME:)
This is amazing. I know that IBM’s Z/OS (formerly OS/390), AIX, and Linux run on mainframes, But I had never heard of any OS from Microsoft running on mainframes. I know that NT used to support DEC Alpha’s, but MS dropped support. And they have a Beta version for the Intel 64-bit, and AMD 64-bit architectures.
If Microsoft is writing an OS for mainframes, this is a major change in their corporate strategy.
Of course it could be that you have no idea what a mainframe is, and are just using the word ‘mainframe’ to mean ‘server’. That’s kind of like calling a honda civic hatchback an 18 wheel Mack truck. You can say the words, but they are complete nonsense.
Arm yourself with knowledge firend and visit the IBM website and learn how different a mainframe is from a micro-computer:
http://www-1.ibm.com/servers/eserver/zseries/990.html
Hint: the reason the kind of computers Microsoft runs on are called micro-computers is because they are much smaller, and less powerful than mainframe computers!
Hey,
I _like_ wolk kernels but…
2.6-test<current bk> is much faster than 2.4.22.
That’s it.
Does anyone know where I can read about these different kernel patches and what the difference is between them? Like -mm4 and -oc or something like that? Also are there any sets of patches you would recommend for the 2.6-test5 vanilla kernel?
Ah well must be that NetBSD is much faster/more efficient then.
Well, I recommend -mm4. Its got improved stability and speed patches.
But for all the info and patch lists:
kerneltrap.org
But nowadays, the vanilla 2.6 kernel is pretty fast.
Startup speed is still the weak-point of KDE. KDE contains some large C++ frameworks, and C++ frameworks are slow to load. Starting Konqueror on my laptop (P4 2GHz, but only 5400 RPM hard-drive) takes about a second. Starting smaller apps (like Konsole) takes only about half a second. Its not instantaneous, like BeOS was on a much slower machine, but its certainly not onerous. It doesn’t bother me much at all, because I almost never start new apps — I just use the multiple desktops feature to keep all the apps I’m working with open at the same time.
“DevFS do work now, but it’s deprecated. It’ll be removed in 2.7.”
On Gentoo i’m using it, and i like it. Anyhow, *why* is it beeing removed? And 2.7 is just development, so it doesn’t really matter much there. For us users, we can still use DevFS for quite a while, taking we’re running stable versions like 2.4 and 2.6 for the next year (probably yearS)
I like WOLK too, i like Con Kolivas patchset even better, but 2.6 still mops the floor with all my 2.4 kernels.
My system feels more interactive, my system boots faster, my system shuts down faster.
It’s al just a bit smoother without huge speed differences (though i think ooo starts faster too, i should benchmark it when i have the time.
some things won’t work like wine, vmware, …. though
I’m also using Gentoo and I have no problem with it… However, I think they want to get rid of it because it’s either a pain to maintain or to develop for. I’ve read many articles where people said the code was a big mess. To what I’ve read, udev is much cleaner and does almost the same thing. I wouldn’t be surprised if Gentoo 1.6 (or 2.0) dump DevFS in favor of udev.
I know it’ll be with us for a few years, but why would you use something that is marked for death?