Re: [Xen-users] Performance of PVM and HVM

["Fajar A. Nugraha" <list@xxxxxxxxx>]

On Mon, Sep 17, 2012 at 12:58 PM, Robin hp <pingh.cd@xxxxxxxxx> wrote:
>>How do you test native?
>
>>For "fair" comparison, you need to make sure that native has the same
>>cpu and memory as dom0/domU tested. For cpu, you can disable some of
>>them using /sys/devices/system/cpu/cpu1/online, and for memory you
>>might need to use "mem=2G" kernel boot parameter.
>
>>Thanks for Fajar's reply.
>
> Yes, All native,pvm and hvm have the same configurations.

Are you sure? setting native parameters can be tricky.

>
>
>>Do you file:/ ? If yes, that would be the effect of dom0 cache. Try
>>using block devices with phy:/
>
>>Others might comment more about this.
>
> I didn't use phy:/ for hvm, but I used file:/ for pvm, and the results were
> not different.

meaning what? you've retested with phy? if not, you should.

> For more testing, I used dd (bs=1024M count=8)to test writing speed for
> native, pvm and hvm. Following is the results:

dd is almost useless for benchmarking purposes.

Personally I'd say use real workloads. For example, in the past I was
choosing the best virtualization platform to run mysql. So I used
sysbench on vmware, kvm, and xen. The end result was almost the same,
so I ended up choosing vmware for that project because it was easier
to maintain.

> What wired me is that most docs about xen told me pvm's performance is
> better, but the results of my testing told difference.

It depends.

64bit pv have some performance penalty for some syscalls which does
not exist in 32bit pv or hvm (search the list archive for details). In
some cases this doesn't matter, and other pv-guest benefits more than
compensate for this penalty.

There's an ongoing effort to create hybrid guest to address this
issue, but it's not ready yet.

Again, at this point I suggest you:
- use phy
- use a benchmark similar to the workload you're going to use

... and if you have time, try kvm as well :)

-- 
Fajar

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users