Mailing List Archive

How can PV be slower than PVHVM?
Hi there,

after reading the PV vs PVHVM performance comparison at

https://xen-orchestra.com/blog/pv-vs-pvhvm-on-next-xenserver/

but I still cannot accept how a fully emulated system can even come
close to the performance of a PV-only system. Every translation layer
must add some overhead - right? So a true PV system without translation
layer should (imho) be at the top of the performance list?

Or does the mentioned comparison apply only to debian on xen-orchestra?

My system is a barebone gentoo xen 4.10 Dom0 and multiple gentoo PV
DomUs in PV mode - and now I'm confused...



_______________________________________________
Xen-users mailing list
Xen-users@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-users
Re: How can PV be slower than PVHVM? [ In reply to ]
Hi Steffen,

This is because (in short and roughly):

1. Linux guest kernel is detecting Xen drivers on boot and then it will
bypass any emulation (hypercalls)
2. Modern hardware is "cabled" to use virt instructions (faster than
software), which is using HVM (especially for memory allocation)

So it's not a "fully emulated system" like you imagine, but instead calls
to Xen and using direct hardware capabilities. This is why it's faster than
PV for majority of the load (in short, almost everything related to memory
translations)

Note this a short explanation without getting deeper into details :)

Best,

Olivier.

On Thu, Jan 17, 2019 at 1:14 PM Steffen Einsle <einsle@phptrix.de> wrote:

> Hi there,
>
> after reading the PV vs PVHVM performance comparison at
>
> https://xen-orchestra.com/blog/pv-vs-pvhvm-on-next-xenserver/
>
> but I still cannot accept how a fully emulated system can even come
> close to the performance of a PV-only system. Every translation layer
> must add some overhead - right? So a true PV system without translation
> layer should (imho) be at the top of the performance list?
>
> Or does the mentioned comparison apply only to debian on xen-orchestra?
>
> My system is a barebone gentoo xen 4.10 Dom0 and multiple gentoo PV
> DomUs in PV mode - and now I'm confused...
>
>
>
> _______________________________________________
> Xen-users mailing list
> Xen-users@lists.xenproject.org
> https://lists.xenproject.org/mailman/listinfo/xen-users
Re: How can PV be slower than PVHVM? [ In reply to ]
On Thu, Jan 17, 2019 at 1:14 PM Steffen Einsle <einsle@phptrix.de> wrote:

> Hi there,
>
> after reading the PV vs PVHVM performance comparison at
>
> https://xen-orchestra.com/blog/pv-vs-pvhvm-on-next-xenserver/
>
> but I still cannot accept how a fully emulated system can even come
> close to the performance of a PV-only system. Every translation layer
> must add some overhead - right? So a true PV system without translation
> layer should (imho) be at the top of the performance list?
>
> Or does the mentioned comparison apply only to debian on xen-orchestra?
>
> My system is a barebone gentoo xen 4.10 Dom0 and multiple gentoo PV
> DomUs in PV mode - and now I'm confused...
>

Especially compiling anything on Gentoo with PVHVM is a solid boost-up
compared to PV.


>
>
> _______________________________________________
> Xen-users mailing list
> Xen-users@lists.xenproject.org
> https://lists.xenproject.org/mailman/listinfo/xen-users
Re: How can PV be slower than PVHVM? [ In reply to ]
Am Donnerstag, 17. Januar 2019, 15:22:22 CET schrieb Tomas Mozes:
> My system is a barebone gentoo xen 4.10 Dom0 and multiple gentoo PV
> DomUs in PV mode - and now I'm confused...
>
> Especially compiling anything on Gentoo with PVHVM is a solid boost-up
> compared to PV.
hmm,

this sounds to me a bit if a suboptimal vCPU setup / Gentoo build environment
is used in PV.

i do a lot of compiling on source based distros like Gentoo, *BSDs etc. and i
hate all the full VM stuff from VMware, KVM/QEMU etc. because it is
significantly slower - especially under any others / parallel VMs load on the
metal. So i'm wondering too....ß)

But did not tried PVHVM yet byself.

Could someone explain this deeper?


just my .02$
best regards,


niels.

--
---
Niels Dettenbach
Syndicat IT & Internet
http://www.syndicat.com
PGP: https://syndicat.com/pub_key.asc
---








_______________________________________________
Xen-users mailing list
Xen-users@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-users
Re: How can PV be slower than PVHVM? [ In reply to ]
Hello,

On Thu, Jan 17, 2019 at 01:10:29PM +0100, Steffen Einsle wrote:
> after reading the PV vs PVHVM performance comparison at
>
> https://xen-orchestra.com/blog/pv-vs-pvhvm-on-next-xenserver/

Note also that this article is from January 2016 and is based on Xen
4.6. The advent of Spectre and Meltdown in my opinion would warrant
a re-run of these benchmarks with the trade-offs selected as
appropriate for your use case.

Cheers,
Andy

_______________________________________________
Xen-users mailing list
Xen-users@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-users