general xen status wrt linode

We're done on our end of things; it's Xen that we're waiting on maturing. I read the other day that the developers consider 3.0.3 beta quality (wha?), so that explains why I've been disappointed – my expectations were too high. Once Xen does mature (and it appears that it is doing so) we'll still have one more roadblock: We need control over disk I/O, more than what just CFQ/ionice provides. I have a few ideas on doing this, but none that I've implemented yet.

-Chris

This makes me sad. Xtrahost and Memset have xen vm's and I'd been pondering switching but hadn't for quite some time. I now realise the reason I haven't switched is because I absolutely love the whole linode.com website and web interface and general community. They don't seem to have any web based control whatsoever. Anyway, it sounds like Xen on linode is a long way away. :(

You have to wonder about the wisdom of building a business using beta quality software. I'm on the Xen beta at Linode and unexpected server reboots are still too frequent for 'production' use. You can get 100 to 300 day uptimes with UML on Linode. Xen is in the 10 to 30 day region - an order of magnitude less. My UML Linode only had to be rebooted for kernel changes or (virtual) hardware upgrades - Xen isn't there yet.

I don't understand why Xen is better for users. Would we be able to do more, or would we get significantly better performance?

@pclissold: True. If the other providers are suffering from the nature of instability you mention I'd be rather alarmed at what they call service. I might get in touch with them and ask them about how they are finding it so far. Of course I can't be sure whether they will be truthful about it.

@astrashe2: Better performance (near native). I'm sure benchmarks are plentiful around the web.

@astrashe2:

I don't understand why Xen is better for users. Would we be able to do more, or would we get significantly better performance?
Xen has better performance for kernel calls. The interesting thing about User Mode Linux is that the overhead is in kernel calls – userspace runs at native speed. From the kernel build time benchmarks I've ran on non VT/Pacifica hardware, Xen is quite faster than UML but still lags behind native speed by about 20%. Its closer to native speed than UML, but it wasn't "native speed" as I've seen Xen advertised/described elsewhere.

UML is getting SMP support, which would essentially (at least) double the potential CPU performance. UML is also getting some kind of VT/Pacifica support (Intel is working on this), and although I'm not exactly sure of details, it's a pretty obvious guess that it'll affect performance.

Also, Linode just took order of 25 VT-enabled machines, and all future machines will support it. Regardless of the the virtualization platform we'll be using there are performance increases in Linode's future besides just throwing faster hardware at the problem.

-Chris