Love-sources - The "Wailing like Bob Dylan" releas

Message

nephros · Post by **nephros** » Thu Oct 09, 2003 12:43 am

On the Athlon errata/prefetch issue:
I have been running a dual AthlonXP 1800+ (the old one, model 6 stepping 2) on a ASUS a7m266-d mobo for about 6 months now, using kernels from 2.4.17 - .21 and 2.5.54-now (the -mm variety of the day), and never had to issue mem=nopentium for it to boot or work.
It is possible (is it?) that the BIOS incorporates a fix, because Alan mentions something like that in the SMP FAQ; for this mobo you have to do mem=nopentium with older BIOSes but not with the new ones.
I dont know anything about the details of this, so just FWIW.

On the issues of Athlons, the patch from this thread,while purely cosmetic, has worked well for me. It was just not nice to have MCA exceptions on every boot. (The thread also freatures a hilarious objection from Linus IMO.)

Lovechild, world domination is, in effect, the only thing worth fighting for, so keep at it and thank you.
I, for one, will welcome a new overlord who bears "love" in his name.

Regor · Post by **Regor** » Thu Oct 09, 2003 10:26 am

Well I couldn't resist giving the love sources a try this time around and it looks good to me. The most notable thing so far is that I now don't have to apply the vesafb patch myself

Anyway, nice work Lovechild, thanks!

Lovechild · Post by **Lovechild** » Thu Oct 09, 2003 12:31 pm

Okay, Nick Piggin has agreed to do a small interview since people have been asking me questions about the internals of his code.. so what do you people want to know specifically?

Evangelion · Post by **Evangelion** » Thu Oct 09, 2003 12:56 pm

Lovechild wrote:Okay, Nick Piggin has agreed to do a small interview since people have been asking me questions about the internals of his code.. so what do you people want to know specifically?

With his patches not getting in to the 2.6, what are his plans? to get them in to 2.7? The scheduler-work has been late in the coming (when it comes to the desktop that is) in the 2.5/2.6, and I for one would like to see more and more fundamental work on this in the 2.7/2.8-series.

GoinBald · Post by **GoinBald** » Thu Oct 09, 2003 12:57 pm

What's the most powerful force in the universe?

Love!
(It's true, it drives one to surpass any limits)
(Sound's kinda like care-bear power. lol)

I'd like to know the recommended schedulers for vairous block devices
and if I may a low down on the them. You've given a overview but there's little that I've been able to find for further reading.

Where's the future heading with this...ACLs with MAC coexistance etc?
What trade-offs would be made for this...performance?

Just off the top -o- me head on a sick early morning,
j

Lovechild · Post by **Lovechild** » Thu Oct 09, 2003 1:16 pm

GoinBald wrote:What's the most powerful force in the universe?

Love!
(It's true, it drives one to surpass any limits)
(Sound's kinda like care-bear power. lol)

I'd like to know the recommended schedulers for vairous block devices
and if I may a low down on the them. You've given a overview but there's little that I've been able to find for further reading.

Where's the future heading with this...ACLs with MAC coexistance etc?
What trade-offs would be made for this...performance?

Just off the top -o- me head on a sick early morning,
j

Well, google I guess is your friend here but basically:

Any device that will not benefit from scheduling technics because of technology - like flash where you read the grid at equal speed, there it would be silly to enforce a read and wait policy fx.

If you have a system with many users I know there's work doing into integrating user control with CFQ so you can limit users IO bandwidth - useful for servers and timeshare systems I guess.

AS scheduling has great potential for interactivity, I would recommend reading RML's OLS paper for this.

Deadline is basically just a CSCAN scheduler with a deadline for requests so we can ensure requests have to executed within a certain time of being queued - it's a good idea, AS is better though.

it all depends on your needs, but devices suchs as flashstorage doesn't benefit at all from scheduling and such be set to noop.

bssteph · Post by **bssteph** » Thu Oct 09, 2003 6:08 pm

I'm interested if there's any chance of shoving Con's work into Nick's work and coming out with some uberscheduler.

I think I saw on LKML once (maybe from Nick himself) that the methods interfered enough so they wouldn't play nice together (both change the same thing in different ways), but maybe Con's could be tuned down a bit... ? Just wondering what the man most qualified would say. :)

Lovechild · Post by **Lovechild** » Thu Oct 09, 2003 7:37 pm

bssteph wrote:I'm interested if there's any chance of shoving Con's work into Nick's work and coming out with some uberscheduler.

I think I saw on LKML once (maybe from Nick himself) that the methods interfered enough so they wouldn't play nice together (both change the same thing in different ways), but maybe Con's could be tuned down a bit... ? Just wondering what the man most qualified would say.

None at all as I see it, since much of Cons work is based on making the interactivity estimator do it's job correctly - and Nick completely removed that algoritm in his scheduler... that's right we don't have the estimator at all.

TheCoop · Post by **TheCoop** » Thu Oct 09, 2003 7:51 pm

-test7 has just come out...

Lovechild · Post by **Lovechild** » Thu Oct 09, 2003 7:52 pm

TheCoop wrote:-test7 has just come out...

I already posted about that the moment I saw it - not love related yet though.

frippz · Post by **frippz** » Fri Oct 10, 2003 3:35 am

Just booted into the new kernel, but I must've missed something cause I get this errormessage when /etc/init.d/local starts at boot:

Code: Select all

/sbin/runscript.sh: line 7: /sys/block/hda/queue/io_scheduler: No such file or directory

Other than that, I'm still haunted by laggy sound under loads (listening to XMMS and loading Mozilla Firebird as an example) and the garbled console when going out of X. *sigh*

I'm using a Terratec DMX XFire 1024 (using the Cirrus Logic CS_46XX compiled into the kernel).

Zeddicus · Post by **Zeddicus** » Fri Oct 10, 2003 5:00 am

Just so ya know, another satisfied lover :p

First dev kernel to get framebuffer working for me. Thanks!

chrisyu · Post by **chrisyu** » Fri Oct 10, 2003 5:07 am

*Levi* wrote:Just booted into the new kernel, but I must've missed something cause I get this errormessage when /etc/init.d/local starts at boot:
Code: Select all
/sbin/runscript.sh: line 7: /sys/block/hda/queue/io_scheduler: No such file or directory

You need

Code: Select all

mkdir /sys

then reboot

frippz · Post by **frippz** » Fri Oct 10, 2003 11:22 am

Ok, got that issue fixed. Thanx alot!

Now, I only have to resolve the sound issue and the garbled console.

I've tried to compile my soundcard as a module in the kernel, but that didn't change anything...

GoinBald · Post by **GoinBald** » Fri Oct 10, 2003 12:23 pm

Mr. Love,
Thanks for the prompt reply.
I 've found a few references for the schedulers.
But I've been unable to find the cfq home to checkout the ACL work for CFQ.

http://kerneltrap.org/node/view/580
and
http://www.cs.rice.edu/~ssiyer/r/antsched/

Once again,
Thank you!!!!
p.s. The ability to select one for this device and another is awesome.
The best part, no rebooting, no messing with the bootloader

Your lookin at one really appreciative noob.

j

Oh ya Levi,
What's the issue with your sound?
A simple way that works for me but YMMV

relink your old 2.4 (scoff,scoff) kernel.
reboot into the 2.4 ker and:
follow the alsa guide for your chip then,

emerge --oneshot --nodeps alsa-driver
emerge alsa-utils
rc-update add alsasound boot

relink your love kernel and reboot into "lovin joy" again.

HTH,
Baldy

474 · Post by **474** » Fri Oct 10, 2003 12:27 pm

I have recently installed -love1 sources, mainly on account of the presence of Nick Piggins' scheduler policy patch and the per-block device scheduler tunables. Prior to this, I had been experimenting with a 2.6.0-test5 kernel with Nick's scheduler policy patch v15.

When I first became aware of Nick's patch, a few comments immediately grabbed my attention (reproduced here from a thread on kerneltrap.org)

Backboost is gone so X really should be at -10 or even higher.

If anyone does try my patch, it is important that X is reniced to -10 or -15. Whether this will starve other things is up to debate, but it really doesn't get enough CPU if its not reniced. Email me testing experiences and CC lkml as well if you're interested in joining in.

These comments were made upon the v12 release, but I assumed that they still held true, despite my reservations about the very idea of nicing X. So, I did try renicing X with my patched -test5 kernel but it seemed to cause poorer performance, on the basis of a casual observation involving scrolling tests and window dragging with a moderate background load.

So I have two questions:

Have there been any adjustments since the v12 patch which would affect the above?
Has anyone attempted to renice X under the v15 patch (including -love1) and if so, would they have any observations on the effect?

frippz · Post by **frippz** » Fri Oct 10, 2003 12:30 pm

GoinBald wrote: Oh ya Levi,
What's the issue with your sound?
A simple way that works for me but YMMV
relink your old 2.4 (scoff,scoff) kernel.
reboot into the 2.4 ker and:
follow the alsa guide for your chip then,

emerge --oneshot --nodeps alsa-driver
emerge alsa-utils
rc-update add alsasound boot

relink your love kernel and reboot into "lovin joy" again.

Well, the only problem is that there's almost no traces of any 2.4-kernel left on my system. By the way, how will that resolve my choppy sound?

Thanx for the reply in any case...!

kerframil:
I'm using a wrapper for X renicing it to -10 and I am using test6-love1 ATM. Although the only experiences I can share is my choppy sound. But that has haunted me every since I startedt to use the 2.6-kernels. My desktop is otherwise very responsive but things can become really sluggish whenever there is load, so it is not just my sound that suffers from high loads.

474 · Post by **474** » Fri Oct 10, 2003 12:41 pm

kerframil:
I'm using a wrapper for X renicing it to -10 and I am using test6-love1 ATM. Although the only experiences I can share is my choppy sound.

Well, if we assume for the moment that everything I mentioned in the previous post is irrelevant, then renicing X may not help because it could steal CPU cycles away from XMMS where it isn't merited.

Other than that, if you are using the ALSA plugin I would say do not use Mmap mode for the plugin. That usually causes all manner of problems in my experience. If that isn't the problem, then try playing with buffer sizes I suppose. And if you're using ACPI then consider trying without, that can cause odd issues with interrupt routing on some setups.

GoinBald · Post by **GoinBald** » Fri Oct 10, 2003 12:53 pm

kerframil,
First thanks for your scripts for openssl 9.7 and distfiles

Yes I've reniced X and since test5-love4 have not noticed anything different. Prior, yes there was a major difference.
Wapping other apps ie xmms,firebird etc. helps.

Levi,
Sry I didn't know you had choppy sound, setting priority helps.
Also you might wanna try powertweak as this will tell you the recommended buffer sizes for your hw.

Just another noob,
Baldy

frippz · Post by **frippz** » Fri Oct 10, 2003 12:54 pm

kerframil wrote: Other than that, if you are using the ALSA plugin I would say do not use Mmap mode for the plugin. That usually causes all manner of problems in my experience. If that isn't the problem, then try playing with buffer sizes I suppose. And if you're using ACPI then consider trying without, that can cause odd issues with interrupt routing on some setups.

I've just disabled Mmap in XMMS and as we speak, I am recompiling the kernel without ACPI (how the heck did it get there in the first place??).
I'm not sure how to fiddle with the buffer sizes, but before I go about something like that, I'll try to reboot with the new kernel and then try to remove the wrapper for X. If there's still no dice, I guess I'll go about the buffers.

Lovechild · Post by **Lovechild** » Fri Oct 10, 2003 3:04 pm

Okay it's about time to start a new thread...

Since I just started assembling test7-love1, the major news would be switching from -mm to -wli to test some VM improvements, other than that it's just a sync with Linus.

So any requests?

-edit-

did I mention how sweet the test7-love1 prerelease feels?

didl · Post by **didl** » Fri Oct 10, 2003 3:11 pm

Lovechild wrote:
did I mention how sweet the test7-love1 prerelease feels?

Can't wait!

test6-love1 is about the best thing that ever happened to my
thinkpad (besides the fact that track pad and track point refuse
to work simultanously)

474 · Post by **474** » Fri Oct 10, 2003 3:16 pm

Well, I can't say I've been overawed by recent -mm patchsets, so this should be interesting. I must confess I had no idea who WLI was until you mentioned it

Nick doesn't appear to have updated his patch for test7 yet so I presume your next patch will leave things as they are (i.e. Con Kolivas' work)? Also, will the tuneables (if I'm not mistaken, those were supplied by Nick) stay in?

As for feature requests, some of the stuff from the OSDL patch branch looks interesting - not that I've investigated in any detail yet. See this link.

Lovechild · Post by **Lovechild** » Fri Oct 10, 2003 3:24 pm

I just have one tiny problem to figure out... my computer started to randomly reboot...

I'm guess I have to find wli and talk to him.

-edit-

ACPI seems to be broken.

474 · Post by **474** » Fri Oct 10, 2003 4:02 pm

ACPI seems to be broken.

Why am I not in the least bit surprised. I would point out that "power-off-on-shutdown" support actually seems to work properly with APM in 2.6, so unless there is a specific need for ACPI, I still subscribe to the policy of avoidance! 2.4 kernels also work well with pci=noacpi as a compromise, but I'm not sure if that is the case in 2.6 at present.

Love-sources - The "Wailing like Bob Dylan" releas

We get to ask the masta?

Fair allocation of resources to X