Hi Falk, hi everybody,

First: I am glad to help you.

At the moment I run a 32Bit singlecored Linux VServer (2.6.30.3-skas3-kernelsendpagefixed) in uid-based cgroup scheduling with 3 concurrent instances of an uml-kernel (with one TUN-network each) and a "share-16" performanceconsuming task (as the idle employment).

Each uml-kernel runs under a seperate user (user 1 shares : 1024 for the IPv6 testrouter(linux 2.6.24), user 2 and 3 share: 512 for a SAMBA testserver and a crosscompile environment(linux 2.30.5)) for about 24h and all 3 seem to be stable. (but the router and the server currently havn't much load)
I use (limited) shm on the host (384MB RAM / 512 MB SWAP), so memory-operations are speed up vast. ("shm on /dev/shm type tmpfs (rw,nosuid,nodev,size=768M)")

For now I know 2 things about them (both uml-linux-versions):
    1) The UML-implemented (in guestsystems) COW-Filesystem won't work correct and produces more and more bad-sectors over timer (so only use real images for each uml)
    2) The "advanced syscall emulation patch for ptrace" seems to be buggy (maybe in host AND guest) and leads to crashes - so I patched the uml-kernels to don't use this
    (perhaps this also can be achieved through uml-commandline?)

In performaces criterias the uml is only compareable with the host-performance, if only one process is active in UML (tested with openssl).
UML-conextswitching seems to cost enormous cpu. (But no problem, the host-cgroup scheduler makes it fair under the 3 UMLs)

Long story short:
UML is a great thing to virtualize your linux, for example if you need safe environments with rootright (compile environment) - without compromising your productive system, to test Software or to host cheap and small servers for small usergroups. It is a great way to decouple and detach systems.

btw.: How do you feel about UML, what bugs had you discovered? - I am thankful about all kind of information.

-Thanks in advance, Stephan Bärwolf




Falk Husemann schrieb:
Hi Stephan,

you just made my day. I was ranting about why Debian lenny didn't contain the skas-patch anymore and there was no recent working patch out. Even started looking at OpenVZ.

So can you let me and the list know if and how stable it works? I'm interested in performance when used with uml_switch and TUN and lots of concurrent instances. Any experience so far?


Greets,
Falk

On Sep 15, 2009, at 12:04 PM, Stephan Bärwolf wrote:

Hi everybody,

Because I needed the SKAS3-Patch for the Linux (Host)Kernel 2.6.30.3, I
patched this version.

So perhaps anybody else would like to take it - here is it.

The patch can be downloaded at
http://www.matrixstorm.com/software/linux/2.6.30.3-skas3.patch, but the
file is also in the attachments.

Sincerly, Stephan Bärwolf - Ilmenau/Thuringia/Germany
This is a SKAS3 patch for Linux 2.6.30.3. If you have any questions or
problems, please ask on the uml-user list (URL below).
-- Stephan Baerwolf <stephan@matrixstorm.com>

The usual maintainer of such patches:
--Ryan Finnie <ryan@finnie.org>

Download location:
http://www.matrixstorm.com/software/linux/2.6.30.3-skas3.patch

uml-user list:
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-user