I'm half-considering something like this for a project I'm working on, actually, so don't laugh too hard! In theory, it should work, and it might also be pretty effective and efficient, but with a few caveats:

1) IP addresses. This might be valid justification for additional IP addresses from Linode, but I'd probably check that out before writing up the business plan. Most customers will want a real, public IP address, even at the $8/mo price-point. However, "World's FIRST fully-IPv6 VPS Service!" would probably get you... a couple customers.

2) Kernels. You'd be doing the pvgrub thing, which would make your situation relatively unsupported. Of course, it's not like there are ever any bugs in the kernel, so why worry?

3) Management. Generally speaking, most OpenVZ hosts are probably running dozens of containers, which means providers can grow a fair bit before having to seriously automate things. In this case, with four customers per host, you're going to have to automate the daylights out of things from Day 1. With just 100 customers, you've got 25 OpenVZ servers while Chet Bumpkin's Live Bait and VPS Shoppe has two. This is not necessarily a bad thing.

But you know, like that time someone popped onto IRC and wondered about running Plan9 on a Linode, all I can say is "you're on your own, but let us know how it works!" -rt (paypal me a cut while you're at it)

I believe that everybody who has attempted it in the past has met with strange issues, but it looks like it works with newer versions of Xen:

http://wiki.openvz.org/How_to_use_OpenV ... 6_platform)

I think that Linode is still on Xen 2.x, though, aren't they? Perhaps it will be possible in the future. It's certainly possible to see the advantages of running a small VPS business at Linode, due to the scalability of the cloud-based service, and the ability to use stack scripts to make deploying additional VPS servers extremely fast.

But still... VPS-within-VPS feels... wrong.

UML!

_________________
Disclaimer: I am no longer employed by Linode; opinions are my own alone.

UML is a lot slower than OpenVZ... Note the big performance increases Linode customers saw when the switch from UML to Xen was made!

I've also had limited experience (ie, dabbling if anything) with virtualization outside of Xen, VMWare, and VirtualBox. However, OpenVZ is faster than Xen due to less virtualization overhead, anecdotal evidence from Linode customers (and various documentation on the net saying as much) seems to indicate UML is much slower than Xen, therefore UML is much slower than OpenVZ.

X > Y
Y > Z
Therefore X > Z

I'm not disagreeing that UML doesn't look like the best option for running on a Linode, only that there may be performance concerns.

Linode should find a way to prevent people from doing VPS inside a VPS, it will hurt the overall performance of the physical machines.

The major Linode performance boost from UML to Xen was, if I recall, due to having multiple processors available. I *think* UML will now do SMP, so it might be worth another look.

Also, as far as performance differences go, anything less than 10% is effectively equal in the real world. I'd consider both Xen and OpenVZ to be within 10% of bare metal performance, with the major differences being unrelated to the system in use. (In other words, a shiny-pants dedicated server on a SAN is going to suck relative to a UML Linode.)

</kerosene>

So, uhh, what did I use my RAM for... well, I think most of it went towards cache, but some went towards buffers.

Aside from recursive virtualization being impossible to prevent (beyond the limitations of the CPU architecture), how could it hurt performance for other users?

Since it's Friday afternoon and I'm feeling punchy, I'll point out that there isn't that much difference between JVM and UML from a nebulous-diagram standpoint.

Mmm JVM == Java Virtual Machine? If that's the one you're referring to then there's a world of difference. End of the day a JVM executes just the bare minimum to run a Java app.

The same app let's say written in C++ running on let's say UML now has an entire OS being emulated just to run the app.

How is that not going to hurt performance? I have some background in the virtualization arena but who doesn't.

I guess someone would need to do some measurements but if we combine the two models, and a sub-user (user leasing a VPS running inside a Linode VPS) wants to run Tomcat or (eek, Glassfish or JBoss) to run that same Java app, but the JVM instance is being executed by the whole UML OS on top of the XEN VM running on the host Linode OS ... virtualization compounded. Yeah those numbers should look great!

What we really need is a Java OS which has its instructions run natively on the host CPU. I like Linux but if the OS and the apps and the 'virtual' containers were all Java and the apps running in the containers were Java; that would be cool. No C, no C++, no Perl, no Python, yeah!