PVM Virtualization Framework Proposed For Linux - Built Atop The KVM Hypervisor

[phoronix]

Phoronix: PVM Virtualization Framework Proposed For Linux - Built Atop The KVM Hypervisor

Ant Group and Alibaba have proposed PVM, the Pagetable Virtual Machine, as a new virtualization framework built upon the Kernel-based Virtual Machine (KVM) hypervisor. PVM does not require hardware-assisted virtualization while working with KVM-enabled software like Kata Containers...

https://www.phoronix.com/news/PVM-Hypervisor-Linux-RFC

[bezirg]

I wish I could understand what's happening here. Seems interesting.

[varikonniemi]

I wish I could understand what's happening here. Seems interesting.

i'm totally in the same boat If this was possible, why was it not done previously? Is this completely new computer science theory!?

There always existed such a dramatic performance gap between HW accelerated virtualization an normal. What i read here is that in most cases no HW acceleration is needed, let's just not be stupid!

[dragorth]

i'm totally in the same boat If this was possible, why was it not done previously? Is this completely new computer science theory!?

There always existed such a dramatic performance gap between HW accelerated virtualization an normal. What i read here is that in most cases no HW acceleration is needed, let's just not be stupid!

Hardware assisted Virtualization came before the Page Table Nesting that is so ubiquitous. That Page Table Hardware exists because of the other Hardware VMs need. They continued on the same track.

Now, it is noticeable that Michael didn't mention speed as one of its strengths, so Hardware VM assistance is probably still faster. This probably gets the same hit that Nested VMs get, a major drop in performance each time the Page Table is swapped. I will have to read the Mail to see if this is mentioned.

This is still cool tech, for its safety if nothing else.

[dragorth]

It's mentioned, but I think I jumped the gun assuming they were using Nested Page Tables. They don't currently handle 5 layer paging, so I am not sure what they are doing there.

Both Michael and the mail mention the shadow paging performance issue, mostly with self modifying code. Think JIT/WebAssembly/DotNet. Anything that isn't compiled directly to machine code, and some machine code is self modifying as well. Long running Java processes are mentioned as running well.

So, assume these are best for long running services, and not interactivity for the user.

[JEBjames]

Michael

Typo "Bit for" should be "But for"

[intelfx]

Phoronix: PVM Virtualization Framework Proposed For Linux - Built Atop The KVM Hypervisor

Ant Group and Alibaba have proposed PVM, the Pagetable Virtual Machine, as a new virtualization framework built upon the Kernel-based Virtual Machine (KVM) hypervisor. PVM does not require hardware-assisted virtualization while working with KVM-enabled software like Kata Containers...

Eh, what? Did they just reinvent shadow page tables (which is the approach that predated hardware-assisted virtualization)?

[binarybanana]

In the early/mid 00s VirtualBox (and VMware, etc.) ran just fine without hardware virtualization. Performance was OK, good enough even to play games if they worked with the limited 3d acceleration. Only kernel mode code had to be emulated, everything else ran natively. This is probably even better.

[olivier]

It's basically "re-inventing" Xen Paravirt mode (PV), which was the first Open Source virtualization solution back in 2003. In a nutshell: get Xen PV code, write it as an abstraction layer for KVM and you are done.

Usually, PV vs HVM mode gives around 20% perf boost for HVM thanks to hardware assisted instruction in the metal. For cases where you don't request too many pages too often, it can be lower (and in pure CPU usage without needing memory paging, it can be similar).