BHyVe - ESXi comparison

Andrea Brancatelli abrancatelli at schema31.it
Tue Jan 28 17:30:28 UTC 2014


Hello Peter,

unfortunately we've been a bit sloppy in tracking the time output because
initially it was just an internal test, thus we don't have the details.

We're setting up a new round of tests we'll run tomorrow and we'll track
user/system/real in a more precise way; I will also publish a graph with
the three stacked piles.

Hyperthreading should hopefully be enable on the host, frankly I didn't
check it out, I will tomorrow.

KVM and QEMU are a bit out of our scope, so we didn't have plans for that.
If I can fine some spare time we'll try.


On Tue, Jan 28, 2014 at 5:10 PM, Peter Grehan <grehan at freebsd.org> wrote:

> Hi Andrea,
>
>
>  We did a very rough comparison betweend BHyVe and VMWare ESXi. Maybe
>> you want to give it a read and let me know if I did write a bunch of
>> sh!t :-)
>>
>
>  Looks good to me :) Thanks for running the tests.
>
>  Would you be able to list the command options you used with bhyve when
> running these tests ?
>
>  What I couldn’t really understand (but that’s something not related
>> to bhyve or VMWare) is how a multiprocessor machine is slower than a
>> singleprocessor machine in doing the compilation… any idea?
>>
>
>  Is hyper-threading enabled on your system ? If not, then with a host only
> having 2 CPUs and a 2 vCPU guest, there isn't as much opportunity to
> overlap host i/o threads with vCPU threads.
>
>  It would be interesting to see your "time" results when running bhyve to
> show %user/%system etc - that may give an indication of how much time is
> spent on 'overhead' CPU usage as opposed to pure vCPU usage.
>
> > 20 VM – 2 CPUs – 2GB RAM
>
>  Interesting result to say the least :)
>
>  I'll try and repro this and see if it's something simple. At first guess
> I'd say it's the classic 'lock-holder-preemption' issue that the ESXi
> scheduler has a lot of smarts to avoid.
>
>  Another interesting test would be Qemu/KVM VMs on Linux to see if it has
> the same issue.
>
> later,
>
> Peter.
>



-- 




*Andrea BrancatelliSchema 31 S.r.l. - Socio UnicoResponsabile ITROMA -
FIRENZE - PALERMO ITALYTel: +39. 06.98.358.472*

*Cell: +39 331.2488468Fax: +39. 055.71.880.466Società del Gruppo SC31
ITALIA*


More information about the freebsd-virtualization mailing list