LWN: Comments on "Xen and the new processors"

Linux does require modification

giraffedata — 2006-05-05T00:36:55+00:00

The whole point of full virtualization is that you run unmodified binaries

It's simpler than that. Even at the source code level, Linux must be modified to run under Linux. The reason you can "just build with ARCH=um" is that someone has made those modifications.

When people speak of the value of a virtualization that lets you run unmodified code, they're concerned about all that code that hasn't been modified yet for that particular kind of virtual machine, and probably never will be. E.g. Windows. If there's only one program you want to run in the virtual machine (e.g. Linux), and there's already a modified version available, you can cross that concern off your list and consider UML, XEN, etc.

By the way, if you don't require a virtual machine that exactly emulates some existing real machine, then the universe of virtualization is much larger than most people see. Any Unix process is a virtual machine. Used to be, you could run only one program on a computer at a time. The invention of operating systems like Unix let you have essentially (virtually) multiple computers based on a single piece of hardware, as long as you were willing to modify your programs for a different environment.

editor forgot UML (again)

proski — 2006-05-04T21:21:29+00:00

Try running "unmodified" Windows that you built with ARCH=um :-)

The whole point of full virtualization is that you run unmodified binaries, whether it's Linux, QNX, OS/2 or CP/M. I heard that Linux compiled to run in Xen would run faster than Linux built to run on raw hardware. So you really care about full virtualizaion if you cannot or don't want to recompile anything.

It's amazing that it took so many years to support IA-32 virtualization. The 16-bit virtual machine support appeared in Intel 80386 in 1986 - twenty years ago.

Xen and the new processors

nix — 2006-05-04T16:31:39+00:00

Using VM86 mode in the guest boot process torpedoes things somewhat for running 32-bit guest OSes on Intel x86-64's, doesn't it?

editor forgot UML (again)

linuxbox — 2006-05-04T14:50:12+00:00

I actually didn't notice that UML wasn't mentioned in the list--and I agree. Any way I measure it, UML with SKAS trails Xen in performance, but the feature set is unique. It's changed the way I develop and test Linux software, particularly kernel code.

editor forgot UML (again)

amtota — 2006-05-04T12:10:49+00:00

Funny how most articles published lately about virtualisation (hot
subject) forget to mention one technology which has been in the kernel for
years, is stable and has features that many others lack (ie: x86 NPTL
support was added in 2.6.16, networking options, honeypots..).

I really wonder why that is.
Sure, UML does not provide a way of migrating live instances from one host
to another (which is cool) - but it does run unmodified Linux instances
provided you build your kernel with ARCH=um. Support for AMD64 is now
pretty good, and you can run x86 guests on an AMD64 hosts. And with
skas0/skas3 mode, it is fast.

Maybe something to do with the fact that the devs just get the job done
but do not have a marketing department pushing out press releases like xen
does... Don't get me wrong, nothing wrong with xen, just too many press
releases per line of code for my liking!

I must admit that the sf.net site needs a bit of work (it is being worked
on), but I've started publishing 'uml for dummies' instructions here:
http://uml.nagafix.co.uk
And there is a nice wiki too:
http://uml.harlowhill.com/

Antoine