Re: [Xen-users] Multiseat workstation with one VM per user

[Gordan Bobic <gordan@xxxxxxxxxx>]

On 2015-01-05 12:08, Luis P. Mendes wrote:
> Hi,
>
> 2015-01-01 2:22 GMT+00:00 Gordan Bobic <gordan@xxxxxxxxxx>:
> > On 2014-12-31 12:38, Luis P. Mendes wrote:
> >
> > > - base system (dom0) as lean as possible, just for Xen
> >
> > Trying to lean things out to a great extent is generally a
> > waste of time. On any recent Linux distribution the base
> > install is sufficiently large that it's losing game and
> > saving a few GB of disk space is not worth the effort.
> >
> > > - one Slackware VM and one Ubuntu VM with direct access to hardware
> > > via PV
> >
> > You need to clarify what exactly you mean by this. Getting
> > hardware passthrough working at all can be hit and miss and
> > is very hardware dependant. There are so many hardware and
> > firmware bugs around that luck is a large factor in hardware
> > selection.
> I'd like to have both VMs in a Paravirtualized mode as in
> http://wiki.xen.org/wiki/Xen_Project_Software_Overview#Xen_Project_Paravirtualization_.28PV.29
> with direct access to dedicated graphics card and usb controller or usb 
> devices.
>
> > > - other VMs for occasional use, which can run in virtualized 
> > > hardware.
> > >
> > > - three fanless graphic cards, for example AMD Radeon 6450. One for
> > > base system (could be a cheaper one), and one dedicated (passthrough)
> > > to Slackware VM, and similar for the third one for the Ubuntu VM.
> > > Iâd be using HDMI as the output interface for the two VMs and VGA
> > > for the base system, in case of necessity.
> > >
> > >  I've read http://wiki.xen.org/wiki/Xen_VGA_Passthrough [1] and
> > > http://wiki.xen.org/wiki/Xen_VGA_Passthrough_Tested_Adapters [2]
> > >
> > > but still would like your opinions, as itâs my first time with Xen
> > > and Iâm not fully aware of all the corners I could face.
> >
> > People's experience with ATI cards is at best mixed. I never got
> > it fully working. Most people find it works OK on the first boot
> > of the VMs, but as soon as you need to reboot VMs things fall
> > apart pretty quickly with cards not being reinitialized properly
> > on a reboot. That's on Windows VMs. With Linux VMs, a lot would
> > depent on how up to the job the radeon driver is. Last I checked,
> > it wasn't.
> I've no experience with this, but always got the impression, from what
> I've read, that nvidia proprietary blob was bettter than AMD ATI's on
> Linux, but that open-source ATI radeon driver was better than
> Nvidia's.
> The reboot problem you mention is something I have to take into 
> consideration.
I have no idea how/if Linux drivers in domU handle GPU passthrough.
It is not a particularly commonly used arrangement.

> > If all you are after is Linux-on-Linux kind of a setup, you would
> > probably be a lot better off with something like LXC, OpenVZ or
> > VServer for separating server tasks.
> But AFAICT, Linux on  Linux should have the same OS and the same
> Kernel (maybe it supports a kernel with a minor update difference).
> As I want to have Slackware and Ubuntu, I don't think I can use LXC or
> OpenVZ for that.
Yes you can. With containers you can run any distro's userspace
in the chroot. The kernel is shared, but that is largely
irrelevant (and a good thing in terms of performance, unless you are
concerned about security related to kernel level exploits).

> > If all you need is a multi-seat workstation, you don't need
> > virtualization at all, you can just configure multiple Xorg
> > instances to access different GPU/keyboard/mouse sets.
> As stated above, as I want to have one system with Slackware and
> another one with Ubuntu, I don't see any other way than to have
> virtualization.
See above. You don't actually need full virtualization for that.

> > > 7. (repetition)  Is NetBSD with its lower power requirements up to
> > > this task?
> >
> > There is no gain. Getting this kind of a setup to work reliably at
> > all on OS-es that are used (and thus debugged) by thousands of people
> > is difficult enough without getting bogged down in OS-es that only
> > a handful of people use in a similar scenario.
> >
> > > In conclusion:
> > >
> > > One workstation, with native disk and graphic card access to each of
> > > the two main VMs running as fast as it they were native.
> >
> > As fast as native? Not going to happen. Fast enough? Sure. I have
> > a triple seat gaming machine that works quite well, but that is
> > very different from what you are proposing above (Nvidia GPUs,
> > Windows guests)
> So, for graphics do you get native performance with the passthrough?
I haven't benchmarked it, but the deterioration is sufficiently
small to be tolerable. After I virtualized I upgraded from a GTX680
to a GTX780Ti to make up any performance drop in gaming.

> But for the rest, are you able to measure how slower are the guest
> systems?
It depends on the workload. With a highly parallel workload
capable of saturating the host, the performance can be quite
dire:

http://www.altechnative.net/2012/08/04/virtual-performance-part-1-vmware/

I do regular performance testing with MySQL under parallel, saturation
level loads and the performance degradation hasn't changed to any
significant extent since that article was written.

Gordan


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