Ok, you have a lot of points there
. Second try, now using a NUC. My aim is to install proxmox on the samsung 980 pro, which is an PCle 4.0 NVMe M.2 SSD.
Then, for storage purposes, I will add another NVMe M.2 SSD drive, possible a 2 TB. And indeed my aim is to pass-through that drive to the VM running Ubuntu + NAS + syncthing. Doesn't that make sense?
I wonder why you are fixated on passing through the disk to the Samba VM? If the virtualization overhead is a concern, that it should apply to all of the I/O, disk and network because every byte consumed by your Samba clients will need to pass through the entire pipeline.
But there I believe you overestimate that overhead. In the very old days, say PDP-11 or so, and before the VAX brought virtual memory to the masses, I/O data blocks had to be copied in and out from kernel space to user space. That was also still the case in the early days of the PC, because between the 8-bit DMA chips on the original IBM-PC and the first bus master ATA and SCSI controllers, data had to be moved via the CPU.
But for a very long time now most disk and network I/O only sets up page table entries and controller registers and then let DMA engines do the minimal amount of transfers: disk data will go to RAM without the CPU being involved and then out to the network, again without being copied with the box and the CPU will only do some PTE manipulations and register setups to inform the devices where things should go.
And these days CPUs have gained the ability to delegate PTEs to VMs, so there is no need to transition to the hypervisor even for the orchestration work, which would still mean some overhead. Long story short, with the proper hypervisor-aware drivers, there shouldn't be much of a difference if you operate a Samba server on the host or in a VM. If you use a container, it's pretty much like running it on the host.
Pass-through of host devices to a VM is typically meant to eliminate virtualization overhead in terms of double-buffering (not happening here), ring-transition or hypercalls (eliminated) which cost CPU time and latency overhead. And in case of a GPU these overheads can still be felt most, but mostly they are hard to manage from two different operating systems at once.
For network and disk devices the limitations of the medium are far more likely to limit performance.
The NUC at least has 2.5GBit Ethernet, but unless your network and clients can keep up, throughput might still be limited to 1Gbit or 100MByte/s, which is slower than even a high-perfomance hard disk these days. To keep up with a SATA SSD, you really need a 5Gbit/s network, cheap PCIe v3 NVMe drives will saturate a 10Gbit Ethernet three times over. And with a PCIe v4 device, anything less than 100Gbit Ethernet won't be able to keep up.
In other words, whether you pass the disk through to the VM or just dedicate most of its capacity to that VM, should not make a noticeable difference But you'll be able to test that once you have the hardware.
Other notes on the hardware: that NUC is probably a good match in terms of power consumption/efficiency, but it's not a "great" deal. Note, that I said "NUC class" not NUC...
The biggest issue with NUCs is that they tend to have only a single full NVMe slot because of the space constraints. So if you insist on two drives, you might have to either use a tall variant which supports a secondary SATA drive, or use something more exotic, like the Phantom Canyon NUC, which btw. is a steal at €450 and throws in an RTX 2060m and an extremely competent cooling solution that never gets noisy.
The other avenue is to go with a Mini-ITX that uses "NUC class" mobile technology and there are some really good deals here from Erying, a chinese company that sells on Aliexpress. I have their G660 with an Alder-Lake i7-12700H which offers 3 full NVMe slots, dual SATA, dual Ethernet and after replacing the factory applied paste with a liquid metal runs at 90 Watts sustained using that Noctua cooler very quietly. I run it with a 10Gbit Ethernet NIC in the "GPU slot", because the Xe iGPU is more than enough for that machine.
But the biggest benefit of these NUCs and "mobile" Mini-ITX boards is that you can limit PL1, PL2 and TAU to whatever suits your efficiency/noise needs and tolerances.
If you want to stay with the NUC, PCIe v4 vs v3 should not be noticeable except on the price per GB, but unfortunately the most logical choice, using a Samsung 970 Evo+ 4TB device isn't available, because they only make them up to 2TB...
And note, that the 2nd M.2 slot in this NUC is only connected via one PCIe v3 lane, so it's limited to SATA speeds!
But even with the 2.5GBit/s network, you can really afford to use the cheapest solid state storage you can find, because that's your bottleneck, pass-through or not.
:fill(white):strip_exif()/i/2005363038.jpeg?f=thumbsmall)
:fill(white):strip_exif()/i/2000952891.jpeg?f=thumbsmall)
:strip_exif()/i/2003829330.png?f=thumbsmall)
