virtualbox.org

Don I know, that was my impression too. But according to info directly from the horse's mouth, the Intel is faster (or has less overhead, can't remember) compared to the virtio-net. Maybe due to optimizations in the Intel, or not enough optimizations in the virtio-net. Again, I haven't done any tests, I'm relying on info from a higher power.

socratis wrote:

Technologov wrote:What are the practical uses for UEFI? I can see only two

Absolutely correct. But you wouldn't believe how many people choose EFI, just because it's newer and they believe that their VM will be 2x faster!

Please remember that VirtualBox only provides (Apple specific?) EFI support, not the newer general UEFI.

socratis wrote:that was my impression too. But according to info directly from the horse's mouth, the Intel is faster

I would need to speak to the horse myself! Like I said, virtio-net is effectively a passthrough: a high level direct API to the networking backend in the VirtualBox host app, and the devs had the complete freedom to design that API for efficiency. Everything else is a chipset emulation on top of a passthrough. It simply isn't possible IMHO for any of the emulations to be faster. That one of the emulations is faster than another (i.e. it has a queue and hence needs fewer transactions for the same bandwidth) I have no doubt.

Btw, when I say faster I really mean more efficient. Actual throughput will of course be dictated by the network a lot of the time.

Martin wrote:Please remember that VirtualBox only provides (Apple specific?) EFI support, not the newer general UEFI.

Who said that? VirtualBox always delivered UEFI, based on code from the tianocore/edk2 projects. Yes, the primary target so far is macOS, so it has quite a few tweaks to make it happy. It's on the todo list to improve the UEFI adaption, including non-volatile (U)EFI variable support. Which is something macOS would also like to use...

Ok, thanks for the correction of my faulty interpretation of probably old partially remembered information.

socratis wrote:Don I know, that was my impression too. But according to info directly from the horse's mouth, the Intel is faster (or has less overhead, can't remember) compared to the virtio-net. Maybe due to optimizations in the Intel, or not enough optimizations in the virtio-net. Again, I haven't done any tests, I'm relying on info from a higher power.

When in doubt, measure.

Like everyone else, I agree it doesn't really make sense for virtio to not be the fastest (not counting the MASSIVE performance regression it's had since sometime after ?5.2.8? in some situations) ... but it isn't.

Very basic testing - just flinging a convenient 1.6GB ISO around over GbE - has times of 21s (~76MB/s) for the Intel NIC, 22s (~73MB/s) for PC-Net III, and 26s (~62MB/s) for virtio. And that's on a system adjusted to avoid the virtio bug (which was running at about 5Mb/s most of the time, small "b").

A few months ago I'd have agreed with socratis outright and said it pretty much has to come down to how much time has gone into optimising the various drivers, e.g. if virtio doesn't use zerocopy buffers for example, and so on. But these days I'm more inclined to think that the virtio driver is simply broken, and has a monstrous stall in it somewhere (resource locking / etc). Whatever the reason though, it's the slowest of all three options by a sizeable margin, here at least (5.2.14).

Interesting. However since that sounds like it is not a V6 regression, in this forum I think we should return to a discussion of the V6 beta.