So we still don't have a list of which "modern" GPUs support vGPU?

Use case: jellyfin/plex transcoding, Simple Windows VM workstation, etc.

@Pierre-Briec , @stormi

I had a look at getting the xe-guest-utilities working on Ipfire v2 now (core 173, the latest version).

Using a new /usr/sbin/xe-linux-distribution script, like suggested here, allows it to detect the ipfire.
I then manually copied the binaries and scripts from the linux tar file, into the folders in Ipfire, since the install script did not seem to handle Ipfire properly.

When starting the daemon using /etc/init.d/xe-linux-distribution, the next problem was that the "action" function does not exist in the /etc/init.d/functions file in Ipfire.
So I just edited the script, replacing the "else" with an "fi" in the if testing where the functions file is sources, so that the locally defined action method is used.
Then the agent started fine.

Then I also saw the issue of the IP address not being reported.
In my setup, there are two reasons for this. One is that Ipfire uses "red0", "green0", "blue0" etc as interface names, which the xe-guest-utilities will not consider.
The other reason is that I do PCI passthrough of 3 network cards to the Ipfire, and hence does not use the "vif" interface/network that XCP-ng makes available to the Ipfire. Althought the "green0" is really on the same network as the "vif" in my setup.

This was using the 7.30 tar file from the XCP ISO, I think.
I then cloned the 7.33 / master version of xe-guest-utilities from github, and used that thereafter.

I manually changed and built the xe-guest-utilities, adding "red", "green", "blue" to the list of interface prefixes that got considered, but it did not help.
I suspect the reason is that these interfaces does not have a /sys/class/net/eth0/device/nodename entry, which contains a reference to the "vif" that XCP-ng knows about, as I understand it.
So
/sys/class/net/eth0/device/nodename exists, but the eth0 is not assigned any IP address, since it is not used by IPfire.
While there is no /sys/class/net/green0/device/nodename entry.

I am not sure who is "creating" this "nodename" entry, but I suspect is it Xen.
And I suspect it is missing, since the green0 interface has no relationsship with the dom0 really.

But then I also got more questions around what is actually meant to be displayed of "network" info in the XOA web UI. Is it only the network between dom0 and domU ? Or ideally all networks defined on domU ? (i.e. red0 and blue0 and orange0 ) ?

And I also think I spotted a bug on the "Stats" page of XOA, since under "Disk throughput", it seems like always "xvda" and "xvdd" is displayed, even if the host only has one disk, "xvda". But that I should report as a bug, if I do not find it as already reported / known.

While playing with this, I also noticed that the management agent version was not properly displayed, i.e. not at all. And this seems to be caused by the the version constants not being replaced while building the daemon.
I am not a go build expert, so I'll investigate it a bit more. But it seems like I'm not the only one with that issue, because the same problem seems to exist with the xe-guest-utilities that are part of Alpine Linux 3.17 distribution.

I do not think that there are that many running Ipfire on XCP-ng/Xen. I've been briefly involved in some pull requests against Ipfire, so I might look at making one for getting the xe-guest-utilities into Ipfire itself, but since the use is not high, I have a doubt it makes much sense.

Thanks for a great tool in XCP-ng, I enjoy using it in my home setup.
Regards
Alf Høgemark

@borzel Does this still work with current versions of the guest tools?

Depends a lot on what you have available. IMHO, I consider the system as disposable, since you can export the metadata, reinstall and import the metadata again. Or backup your VM, reinstall, restore. Or migrate out to another host, reinstall, re-migrate the VMs.

So many possibilities, that all depends on your setup.

That's the right way to do it in the future I suppose, however if it's not done, I can suspect it was during early "Citrix times" and decided it wasn't easy to do with the SMAPIv1 structure.

Hopefully, since SMAPIv3 is splitting the things in a better way, it might be easier to make it real 🙂

@olivierlambert never done it myself, but this is indeed exactly what the feature "Copy host BIOS strings to VM" was intended for as @Andrew mentioned. Hopefully the BIOS strings this feature copies are enough for the ROK installer to recognize the "authorized" dell hardware

@poltushima There technically is no CentOS 9, only CentOS stream. CentOS died a premature death when they pulled support for version 8 in 2021. We moved all of our workloads to rocky Linux and haven't looked back.

Stream is basically a rolling beta and not something I would use unless you're just mucking around with it.

Very cool ! Will a great move relying on a single entrypoint.

@rjt You can customize the hook timeout by adding this setting in your xo-server's config:

[xapiOptions] # Timeout in milliseconds syncHookTimeout = 300e3 # 5 minutes

This is unconditional for a reason. The CSTATE errata in Nehalem are crippling - IIRC a core going in an out of a deep C state does not maintain cache coherency correctly, resulting in arbitrary memory corruption.

You really do care about not hitting this errata, even on a test/hobby server.

@splastunov said in Host CPU Statistics:

You can get it with such command

That command doesn't give me the aggregate, it only gives me the CPU for Dom0. I'd like to get that aggregate number.

@jbamford So yes no dynamic memory there. Perhjaps an issue with the Citrix guest tools but never seen this reported.

Hi,

If someone would build this, it should use Xen Orchestra API, as the central orchestrator, that would make most sense to do it that way 🙂

But yes, if someone want to start working on this, we'll be happy to assist on the API part 🙂

As you can see, the first alpha was not released in September 😉

Another XCP-ng pre-release will be done with vTPM, since there's more progress on upstream side (Citrix). When is hard to tell since it's not our call for this component 🙂

We always try to work with Citrix ("XenServer" division now), to push things upstream and manage to get it merged.

@olivierlambert For now I am going to script it with 'xe vm-export" command. Also it would be good to be able to export to the server itself, instead of downloading to the client. My VMs are over 500Gigs

First alpha next week probably. Also consider 8.3 as a base for new stuff, not directly embedding a lot of new things for the official release 🙂 But since it won't be an LTS, it will be easier to make changes there!

@reinvtv said in NFS nconnect support:

As far as I understand it, this will give us multiple tcp streams over an LACP link, truely aggregating traffic on multiple interfaces. (Until now, you needed to use iSCSI multipathing for this, which isnt able to thin provision.).

You're neigher right nor wrong: LACP is more complex than just bundling NICs and in many cases it will NOT give you any benefit.
Why? Simple: For that you need to have NFS using multiple ports but stay with same MAC and IP, which means that your LACP balancing algo needs to be set (and support) L4.
Many just go L2 or L3 (means they decide on MAC or IP-Address).
As a quick search didn't answer me: If it even doesn't go that 'passive range' multiple port way: In typical environments it won't help you at all with LACP.

One might correct my post.

Also interesting: https://vastdata.com/blog/meet-your-need-for-speed-with-nfs/