Re: [Xen-devel] Custom build the Xen0 kernel && miscellany

[Brian Wolfe <brianw@xxxxxxxxxxxx>]

Personally I found it a bit tough to manage configs with the curent
kernel building phase of the xen sources. As a result I wrote the
beginings of a kernel builder for xenlinux images and modules.

If youi hit http://www.terrabox.com/debian/binary you shouyld see a
package called xenlinuxbuilder. It's a simplified method for specificly
building xenlinux kernels based on the installed version of
xen-kernel-packages. To build a specific configuration of a kernel I do:
debian/rules clean menuconfig tarball PRIV=[yes|no]
LIN_VER=[2.4.27|2.6.8] EXTRA=[whatever extraversion label]

note: you MUST specify clean prior to tarball curently IF you apply any
3rd party patches to the source. It doens't handle detecting extra
patches, yet. My goal is to create a mak-kpkg like system specificly for
debian initially, later for general use. It will save your individual
configs into configs/ for reuse after you have fully configured the
individual kernel.

 It will produce a single tarball for /boot and /lib/modules for
privileged kernels, and separate /boot and /lib/modules tarballs for
unpriv kernels. This has made it a bit easier for myself to generate and
maintain separate and easilly built kernels for my various xenlinux
images.

If you are managing more than just 1 or 2 kernel versions then this may
be a smoother method for compiling and packaging since *eventually* it
will produce .deb packages that will verify which xen version it is that
you are running on the underlying xen kernel.

I have also packaged up the xen distro as pre-compiled binaries for xen
on Debian Sarge. I'm close to having a nightly export run that will bk
update, debian patch, compile, and package them automagically. :) I'm
willing to share my custom makefile for doing this if you want to
maintain a local repository of xen packages.

I stil have a good bit of work to do before I can send the debian/* tree
upstream for inclusion into xen proper if the gang will accept them. :)

I also need ideas on how best to tie the debs into the netbsd vhost disk
image so that I can build netbsd server images dynamicly. Right now i'm
taking a look at the  xenoboot software to see if I can't package it up
for debian to keep from reproducing a lot of the work that has already
been done.


as far as number of images that I run per machine, a friend of mine runs
~20 images on a 4GB dual p4 machine at his offices. I run 6 athlon xp
2000+ machines, each with 1GB ram that run from 2 to 10 images each. So
far not a single beta client has complained of any kind of CPU or
performance issues. So far it appears that memory is more of an issue
here than anything else due to the amazingly low overhead of xen when
using NAS or SAN based storage. :)

As far as selling vhosts using xenlinux kernels, I wouldn't make it an
official product just yet. It's stable for local disks, and in general,
however I have run into various issues that have poped up related to
iSCSI and other momentary glitches as I do semi-daily updates of the
cluster. I would let the xen team get to an official 2.0 release before
truly pushing vhosts out as a paid-service (unless you get a few clients
to sign up for them with the understanding that they aren't guaranteed
stable yet).

If you go lok at http://xen.terrabox.com there is a page for companies
to list themselves on that offer xen based virtual hosts that I started.

Hope this helps from the business perspective for you. :)

Brian

On Mon, 2004-09-20 at 07:16, Mark A. Williamson wrote:
> > I'm new to the list.
> 
> Welcome on board :-)  That's another country to add to the list!
> 
> > Can I tell the default build system to use a modified .config file for the
> > kernel build? From the looks of the makefile it appears that much more than
> > patching and compiling is going on with the kernel, so if I do a make
> > menuconfig and make my changes, am I able to install the kernel in the
> > normal way?
> 
> As Keir said, you should put your XenLinux config files into install/boot/ 
> and 
> they will be used when you rebuild.  You could even stick your old Linux 
> configs into there (with the appropriate names, to match the Xen build 
> system's expectations) and they should work (though the Linux build system 
> might have to confirm a few Xen-specific options interactively).
> 
> > The support I'm adding is for SMP, hypterthreading, big memory, and ATI
> > framebuffer support. I'd also like to pull out various bits and pieces --
> > mostly drivers for hardware I don't have, as this would speed up the build
> > process.
> 
> UP guests (that includes dom0) only at the moment.  You'll use both CPUs by 
> running (multiple) uniprocessor domains on each CLU. SMP guests will be a 
> feature of a future version (perhaps a 3.0 release???).
> 
> This also means guests won't see the hyperthreading but instead you can run a 
> domain on each hyperthread - you may want to see if this is better or worse 
> for performance than disabling HT - it'll depend somewhat in your workload.
> 
> > Thank you for your time. I've been watching Xen with some interest for a
> > while now and I'm thrilled that my employer has given me the go ahead to
> > put together this cluster. I would like to feed the results back to the Xen
> > project somewhere, especially where it pertains to performance of virtual
> > web and database servers. It seems that the wiki has gone west (been dead
> > for a while now), so perhaps it would be appropriate for me to submit stuff
> > back to this list for archival purposes.
> 
> I update the Wiki every so often but it doesn't see that much activity at the 
> moment (maybe more after the 2.0 release...).  Contributions to the user 
> guide are useful, too ;-)  You may like to look at it, although it's not 
> fully up to date at the moment.
> 
> > One final question (and this is bound to be unpopular). I would like to
> > have some idea as to how many virtual web servers I might be able to run on
> > dual Xeon 2.8GHz machines with 4-12GB RAM. Our sites are typically LAMP,
> > with medium sized databases. On average, our current web servers are 900MHz
> > PIII machines that don't often get pushed very hard. Perhaps just some
> > anecdotal accounts would satisfy my curiosity. I realize I'm not being too
> > practical here -- my apologies! ;o).
> 
> Not sure about actual numbers, I'm afraid.  Since Xen has a low overhead you 
> should get a large proportion of the available power for running virtual 
> machines...  You'll also get some "statistical multiplexing" benefit, since 
> slack time from idle domains can be used by active ones, rather than wasted 
> (as in the case of a dedicated machine).  The scheduler should ensure fair 
> sharing when multiple domains are busy.
> 
> Cheers,
> Mark
> 
> -----
> September 19th - Talk Like A Pirate Day
> http://www.talklikeapirate.com
> 
> 
> -------------------------------------------------------
> This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170
> Project Admins to receive an Apple iPod Mini FREE for your judgement on
> who ports your project to Linux PPC the best. Sponsored by IBM.
> Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxxxx
> https://lists.sourceforge.net/lists/listinfo/xen-devel



-------------------------------------------------------
This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170
Project Admins to receive an Apple iPod Mini FREE for your judgement on
who ports your project to Linux PPC the best. Sponsored by IBM.
Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.sourceforge.net/lists/listinfo/xen-devel