UML:

$ wget http://www.kernel.org/pub/linux/kernel/v2.6/linux-2.6.16.tar.bz2
$ tar xfj linux-2.6.16.tar.bz2
$ cd linux-2.6.16
$ make defconfig
$ time make -j2 # (Single processor UML)
...
real    15m36.277s
user    11m4.890s
sys     4m22.970s

Xen:

$ wget http://www.kernel.org/pub/linux/kernel/v2.6/linux-2.6.16.tar.bz2
$ tar xfj linux-2.6.16.tar.bz2
$ cd linux-2.6.16
$ make defconfig
$ time make -j6 # (SMP Linode)
...
real    3m8.766s
user    8m27.320s
sys     1m15.980s

So basically a Xenode can compile a kernel in 3 minutes whereas UML takes about 15. Granted, the Xen host is faster, but even adding a penalty of say, 30%, that still is only a 4 minute kernel compile on Xen.

Not bad, eh?

-Chris

It's certainly an impressive result, but the tests were not the same.
Can we see the result with -j2 on both machines?

Best would be -j1 on both as it seems the Xen linode could allocate more than one VCPU (ie run on multiple CPUs). -j1 would suppress much of the SMP versus UP advantage.

UML

# make distclean; make defconfig; time make -j1
..
real    17m29.744s
user    11m53.640s
sys     4m54.100s

Xen

# make distclean; make defconfig; time make -j1
..
real    5m40.046s
user    4m45.270s
sys     0m38.640s

-Chris

Impressive!

Would it be difficult to run that same test on an OS running on the metal? Might be interesting just for comparison.

Can't wait for Xen to be available for everybody!

Same machine, booted into:
Linux host56.linode.com 2.6.15-1-k7-smp #2 SMP Mon Mar 6 15:50:26 UTC 2006 i686 GNU/Linux

make -j1
real    4m43.388s
user    4m27.114s
sys     0m21.094s

make -j2
real    2m27.666s
user    4m28.531s
sys     0m20.854s

make -j6
real    2m27.219s
user    4m29.766s
sys     0m21.260s

-Chris

So, by my calculations, you lose 73% of your performance with UML, but only take a 17% hit with Xen. Dang!