m00nman

Unraid 6.12.14 Cpu was idle besides this iperf3 test

Here you go: virtio: virtio-net cores aren't pinned with virtio like with virtio-net. iperf3 server is on Windows 2025 VM. Also note that it makes no difference in terms of throughput with my host CPU and with my network (2.5Gbit). It did make a difference when I had i7-4790 (non-k) as my host CPU.

Sorry, but because of additional CPU load and extra emulation I find that unlikely.

Can we get core number limit for a VM without pinning the cores for a VM please? i.e. let the unraid kernel decide which cores are to be used for a VM depending on the load instead of permanently assigning cores to a VM?

You've enabled MSI interrupts for devices that only show "LineBased". If 3090 didn't have it enabled by default, that's the only benefit. I had a different nVidia GPU passed through (1660S ?) and it had MSI enabled by default without me changing anything with new nvidia drivers. YMMV on this though, it might still be useful for the GPUs.

This can help, when you have a dedicated GPU in passthrough mode. You are not using it correctly though (although no harm done either). You can enable MSI interrupts for devices where MSI or MSI-X is listed under supported modes. If "LineBased" is the only option under supported modes you did nothing. On every VM I checked MSI was enabled for all devices that supported it. I don't believe it's as big of a deal before as it used to be before for passthrough GPUs as vendors now support passthrough to VM for the most part and they didn't before.

That's a bad combination. Video driver is emulated + VM decodes the video then reencodes it back to send over the internet via remote desktop app. Remote desktop application aren't really designed for streaming videos at full fps, besides game streaming apps like parsec, but then again you are putting a lot if stress on that cpu without a dedicated gpu to decode and then encode a video stream in real time. Can't you stream video directly to your PC?

Pinning is only necessary when you want to have predictable performance for a VM/container, especially under heavy load, at the expense of pinned cores just idling and doing nothing when the VM pinned to them is doing nothing. Lettings the host's kernel distribute the load (no pinning) between unused or underutilized cores will give you better efficiency, and possibly better performance for all the VMs/containers at the expense of 0 predictability whether that particular VM that you use for, let's say, playing games will have adequate CPU utilization headroom for playing games (there are ways to assign priorities for VMs, but it's beyond the scope here). So it really depends on your use case. Enterprise solutions almost never use cpu pinning because they want to extract maximum performance for all of the VMs/containers.

Interesting.... Some thing just happened to me as I upgraded to B760 motherboard from c23x board (haswell/broadwel generation). It took me by surprise and as1064 controller just stopped being recognised. I actually thought I got a lemon motherboard at first, but started googling and found there was a f/w update for many 106x chips with PCIe gen2. But 1064 is PCIe gen3 so it must be incompatible. Well, thanks to this thread I flashed a f/w that was meant for asmedia 1166 on 1064 and it actually worked. I had nothing to lose so... Anyway, thanks OP!

Just to confirm what @trurl said. I just bought 2x Yottamaster 10Gbit/s 5 bay enclosures thinking I could reduce power usage by using the enclosures + a modent laptop. I reduced idle power usage by around 40W, however the issue with the enclosures was that it was contstantly substituting enclosure serial number for some drives' serial numbers at random every boot. So the serial number for drives would change randomly ever boot which caused unraid to think the drives are missing. After messing around for 2 days I returned the enclosures today. The speed was great. If UNRAID used PARTUUID instead of identifying the disks by-id it would likely have worked, but unraid does it by-id so it's a no go, unfortunately. I guess PARTUUID is just a partition id, so it may not be feasable.

I am saying I don't believe Windows XP supports Hyper-V Enlightenments. So for XP, you would need to enable HPET timer. See below. I striked through what you do NOT need to add for XP, but change the HPET line to "yes"

XP may need HPET. I don't believe it supports anything better than that.

First try to change the machine type to Q35. i440fx does not support PCIe, just regular PCI. You need PCIe since you are passing through NVMe drive, which is PCIe If that makes no difference, try changing "threads" to only 1 and bump number of cores instead.

Hi SimonF, I believe it was version 6.11.x, the version that was current at the time of writing OP. Yes, it is my understanding that that you need all of the following options <vpindex state='on'/> <synic state='on'/> <stimer state='on'/>, additionally some people reported settings migratable='off' resolved issues with GPU PCIe passthrough stutter. migratable='on' is only useful when miograting VM from one host to another, and unraid does not support it anyway so it's safe to turn it off for everyone.

Thanks, this will probably help a few people on here who have GPUs in PCIe passthrough. I didn't really notice any difference for my server windows VMs, however, since unRaid doesn't support migration of a VM to different host (like, for example, Proxmox) it makes sense to disable migration by default as well. I added your suggestion to OP. @Lebowski89 have a look at the post above, it may be the fix you are looking for.