|
Kernel Summit 2005: Virtualization
[Posted July 19, 2005 by corbet]
Rik van Riel and Chris Wright discussed virtualization. This is currently
a hot topic, and several OLS sessions will be dedicated to it in the next
few days. At the kernel level, however, it seems that there are not a
whole lot of issues which need to be resolved.
Virtual hosts may have multiple virtual processors; they will schedule
processes between them. The physical host may also have multiple
processors, and it will be performing its own scheduling. Since the two
levels know little about each other, scheduling imbalances can result. Xen
does a certain amount of "rotating" processes around to deal with this
problem. Despite being discussed for a while, this issue does not appear
to be particularly serious.
It was pointed out that the various virtualization implementations (Xen and
user-mode Linux in particular) have their own virtual buses, virtual
drivers, etc. Might there be some benefit in merging them? Perhaps, but
the amount of code involved is quite small.
Merging Xen. The Xen patches create a completely new architecture for the
virtual machine. There have been objections to this approach; it looked
like a maintenance problem, especially as Xen is ported to more real
architectures. So the patches are being reworked, and the
arch/xen directory is going away. Stuff which is truly specific
to Xen will find its way into the drivers or host architecture
directories. With these changes made, opposition to the merging of Xen
should be much reduced.
Linus had to ask: is anybody actually using Xen? The biggest users are, as
expected, in the virtual hosting business. Most of them are still in
relatively early evaluation stages - Xen is a young technology. Xen is
also heavily used by people who want to play with multiple distributions or
otherwise need sandbox systems to work with.
(Log in to post comments)
Did anybody talk about the Vserver project? (it's similar to BSD'ish jails), except with several advantages IMHO.
I believe Rik Van Riel, has been involved in the project at one point.. and I sure would like it in the kernel ;)
http://linux-vserver.org/
I'm currently using User-Mode Linux as a general purpose server, hosted by Bytemark. I have also used UML experimentally. I am interested in using virtualisation technology to support student project work, so that students can learn by having root access to a machine which can be setup and destroyed on the fly, without this interfering with other uses of the physical host. User-mode Linux is currently easier to set up for this kind of purpose, but Xen promises much better performance, particularly when new silicon proposed by Intel and AMD becomes available, with specific support for virtualisation. Running a set of servers, e.g. apache, sendmail, mysql in different VMs offers better sandboxing, as a security failure in one VM should be less likely to result in knock-on security breaches elsewhere, as would occur in a conventional host running multiple software servers.
So I think Xen is likely to become a very significant development, for a number of reasons, and the sooner it has mainstream kernel support, the easier it will be for potential users to get started with it.
The last sentence in the first paragraph has a typo:
"At the kernel level, however, it seems that there are not a whole of issues which need to be resolved."
I think it should read:
"At the kernel level, however, it seems that there are not a whole lot of issues which need to be resolved."
We currently use VmWare for virtualizing a number of internal servers; our intraweb, bugzilla, etc. VmWare is quite nice, but the host machine gets loaded pretty heavily with only few VM's running, and VmWare does cost real cash. Xen would be really nice here.
In fact, I've even considered experimenting with Solaris/X86, since it has (apparently) decent vitualization support, although it does seem a bit complex to manage.
So, while we don't use Xen yet as we don't consider it mature enough, we certainly will give it a hard look once 3.0 hits the streets.
We're experimenting with Xen for use in Grid computing where we need to
build a gatekeeper machine in front of various HPC clusters to run the
various
versions of Globus that the different grid projects need, along with
development versions. I think we're up to 7 at the moment, assuming
there's just a single cluster behind it.
Using Xen means we should be able to run this all on a single box, and so
far it's looking pretty solid (i.e. it works and we've not seen a crash,
asides from when someone tried something a little brave with LVM
snapshots, but that may not have been Xen related).
All good fun!
Chris
Support for many simultaneous users on ONE pc NOW.
Is only needed the capability of give exclusive access to selected monitor, keyboard, and mouse to a given virtual machine. And obvious "need to have" feature, that it does not have.
and I want it on Unbuntu, please.
|
|