Better sequential performances with emulation vs PV tools

[Olivier Lambert <lambert.olivier@xxxxxxxxx>]

Hi there!

Using XCP-ng 8.1 (Open Source fork of Citrix Hypervisor 8.1), testing on Windows 10 or Windows Server 2019. Those guests are in UEFI, so QEMU emulated disks are presented as NVMe. Storage is a very fast NFS server (using NVMe drivers), in a 10G network.

Doing a benchmark with CrystaliskMark 7.0.0, here is the results without any tools installed:

https://i.imgur.com/7ulGKfY.png

And with PV tools (XCP-ng 8.x ones, or even latest Citrix ones, 9.x, same results):

https://i.imgur.com/y2KDDl5.png

As you can see, SEQ1M (sequential read and write with 1M blocks) is almost 2 times **faster** with the emulated NVMe drives than Xen PV drivers. Which is really weird.

However, interestingly, in RND4K (rand read and write with 4K blocks), this time Xen PV drivers are around 3 times faster. That's interesting because it tells us a lot of things. Might be because the Windows NVMe driver is more efficient for large blocks in a sequential order. That might be the sign there's some room for improvement on Windows PV drivers in this case.

Sadly, I have a huge lack of knowledge in this area, that's why I'm asking here :)

Regards,

Olivier (XCP-ng Project leader)