Virtualization in Linux: A Review of Four Software Choices (Techthrob.com) [LWN.net]

Virtualization in Linux: A Review of Four Software Choices (Techthrob.com)

Posted Feb 12, 2008 13:06 UTC (Tue) by Felix_the_Mac (subscriber, #32242)
In reply to: Virtualization in Linux: A Review of Four Software Choices (Techthrob.com) by rvfh
Parent article: Virtualization in Linux: A Review of Four Software Choices (Techthrob.com)


I have checked with the reviewer and the machine he was testing on does not have VT. Therefore
this is not a review of KVM speed.

(Log in to post comments)

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 13, 2008 1:54 UTC (Wed) by xoddam (subscriber, #2322) [Link]

... and for the same reason not a reflection of the performance of *any* of the virtualisation
environments on a VT-capable processor.  Performance of *all* of them is likely to be better
with fully-virtualised hardware than with paravirtualisation that traps everywhere.

Mind you, there are still a lot of non-VT processors out there, so these 'gut feeling'
performance benchmarks are of some use to those of us still running them.

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 14, 2008 3:38 UTC (Thu) by walken (subscriber, #7089) [Link]

Full disclosure: I'm a VMware employee.

VMware would not use VT even if the box supported it, because current VT hardware is slower
than the emulation techniques we can do in software.

VT holds some promise of getting faster in future CPU generations, and it does make things
simpler if you don't already have the software emulation techniques implemented, but it is not
a silver bullet as of today...

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 14, 2008 4:06 UTC (Thu) by xoddam (subscriber, #2322) [Link]

> current VT hardware is slower than the emulation techniques we can do in software.

I'm intrigued (though not *quite* incredulous).

... I suppose that's not the sort of comment your employer would encourage you to elaborate
upon in a public forum, though?

I would have thought the best you could do in a generic way would be to trap, though I suppose
numerous specific performance optimisations are applicable if you're intimate with your
various guest OSes.

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 14, 2008 8:37 UTC (Thu) by walken (subscriber, #7089) [Link]

Well I can't tell *all* about the secret sauce but the main ingredients are already known...
See http://portal.acm.org/citation.cfm?id=1168857.1168860 for example for a relevant published
paper.

VMware can use binary translation - a technique that is actually quite similar to what qemu or
many java JIT environments are doing. It's complex but surprisingly inexpensive if you do it
right. Since we don't directly execute the guest kernel code but rather a translation of it,
we have an amount of flexibility not available in hardware - if we find out that our
translation for something traps too much, we can look at it and figure out an alternative
translation that will trap less.

The first generation VT hardware, on the other hand, is actually quite limited, it provides
all the traps you need to write a trap-based virtualization monitor, but in the end you still
have to handle all these traps. Most notably the page table virtualization is left out as an
exercise to the monitor implementor - this ends up being a lot of traps and with not much one
can do about them other than waiting for future generations of VT hardware.

Don't get me wrong though, hardware virtualization *will* get faster eventually... it just
takes a while though, due to cpu design cycles being longer than software design cycles :)

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 14, 2008 15:35 UTC (Thu) by sg7jimr (guest, #22837) [Link]

This statement about VMWare not using VT is not completely accurate.  In order to have 64 bit
virtual machines an Intel processor must have VT extensions for VMWare to allow it to run.
That implies that VMWare does in fact make use of VT to some extent.  It was not only necessary
for the processor to support VT but those extensions had to be enabled in the BIOS, so it
seems a real dependency and not some marketing thing.

No VT? No fair to *any* of the platforms reviewed.

Posted Feb 14, 2008 14:58 UTC (Thu) by danpb (subscriber, #4831) [Link]

The first generation of VT enabled hardware is not really focused on performance - it is more
about enablement - ie it allows you to run unmodified guests, with the minimal software
tricks. There are many ways to virtualize without hardware support and many of them will
outperform current VT hardware. eg paravirtualized Xen will easily outperform VT enabled
fullyvirtualized Xen. 2nd generation VT enabled hardware is starting to add some interesting
features (NPT on AMD, EPT on Intel) that can make a huge difference to performance bringing
fullyvirtualized guests closer to being on a par with paravirtualized ones. You still need to
have paravirtualized drivers inside your fullyvirtualized guests though - without it I/O will
be terrible.

The article in question is really quite badly done. Plain QEMU is not really interesting
unless you need to do cross-architecture virtualization, eg PPC on x86 - no one will claim you
should use it for native x86 on x86 because emulation is simply never going to be fast enough.

KVM's future is very bright indeed, but if there's no VT then it reduces to just QEMU
emulation, so the review was rather worthless wrt to KVM. If you're talking state of art, then
KVM should be compared against Xen paravirt and Xen fullyvirt since that's the most commonly
used open source virt out there today, and for all its flaws Xen (eg not being in upstream
LKML) does work pretty well in terms of performance.

Finally, paravirtualized drivers should be used in all tests, since I/O performance will
inevitably suck without it and anyone seriously deploying virtualization in production will
use paravirtualized drivers.