[ltp] Re: Applying undervolting patches

[Andrew Barr]

On Tue, 2006-08-22 at 21:31 -0400, David Abrahams wrote:
> Missing context: The following is from Scott J. Henson.

Oops. Sorry.

> Well, for me switching kernels has been nightmarish.  I'm surely
> missing something very simple, but my wireless card has disappeared,
> adn I can't seem to keep the proprietary ATI driver (8.28) working
> under both the old and the new kernels.

Hardware disappearing is because you don't have a driver for it. You
need to find out what the driver is that your wireless card uses and
recompile it for the new kernel (or enable in the configuration it if it
is in-tree).

I don't know about the ATI driver. Out of tree modules, proprietary or
not, that use kernel APIs extensively (like device drivers) tend to
break between kernel versions easily. Many times looking at communities
(forums, mailing lists, wikis) that have a large number of users of a
particular third-party module will have a patch to update them for new
kernel versions.

> > It can be, depending on how low you can go with your
> > undervolting. In my case, at 600 MHz (lowest frequency) I was able
> > to go all the way down to 700 mV, which is the lowest you can go.
> 
> Are you talking about substantial gains due to undervolting, or due to
> a kernel version upgrade?  Scott was suggesting that the kernel
> version wouldn't make much impact.

Well, I was originally talking about undervolting but from what I've
heard both may have an impact. The 2.6.17 kernel has the best power
management infrastructure of any released kernel to date.

> BTW, how do you decide how low you can drive the voltages?

Two steps: find how low you can go without locking up (there's a script
to help you find this out at ThinkWiki), then download a computational
test program (mprime95, also linked to at ThinkWiki) and see if your
hardware is erring in computation of prime numbers. If it is, bump the
voltages up a step (there is one voltage in a step of 16 mV for each
available clock speed) and try again. Repeat this until you can run the
test for a while (a few minutes is what I tried, it fails fairly readily
if there's a problem IME) without errors.

The particular procedure--including how to use mprime95 to test
computation integrity--is at ThinkWiki. Gentoo-wiki.com also has a page
about Pentium M undervolting which includes user-contributed voltage
tables (incidentally, this is where I found a set that worked for me)


> Well, they *were* tested against edgy according to the linux-phc
> site.  But right now I'm conservatively using the dapper kernel with
> patches.

I can't imagine Ubuntu messes with the cpufreq drivers too much so they
will probably apply to both vanilla trees and Ubuntu distro trees of
supported versions.

> > You might have better luck too using Debian linux-image and
> > linux-source packages, even on Ubuntu, as they are very close to
> > upstream.
> 
> Why would I have better luck with that?

They are very close to kernel.org releases, but built using Debian
kernel packaging tools. If you are having better luck with Ubuntu distro
kernels, however, there's no reason to use Debian's packages.

> I have a vague memory that odd and even xx numbers have different
> meanings (?)

You're thinking of the 2.4 vs. 2.5 kernel series. 2.5 was an unstable
development tree that eventually became the current 2.6 releases. There
is no 2.7 for now (nor for the forseeable future due to change in the
development model). The xx numbers have no odd/even significance.

> So far my impression is that although there are ways to bend Linux to
> your will if you have deep expertise, the system as a whole is
> entirely too fragile.  Pull this lever over here, and something over
> there breaks... :(

Well, don't let that discourage you. :) It takes some time to get to
know how to bend the system without breaking it. You learn a lot along
the way. If you need a Windows partition to keep you productive while
you spend your free time on Linux, so be it. It's a great feeling to be
able to finally save your personal files and delete that NTFS partition
once and for all, but it does take time to get there.

-- 
Andrew Barr | http://www.oakcourt.dyndns.org/~andrew/

All parts should go together without forcing. You must remember that
the parts you are reassembling were disassembled by you. Therefore, if
you can't get them together again, there must be a reason. By all
means, do not use a hammer.
  -- IBM maintenance manual (1925)