the last rewrite of the stream processing ( spring 2025 ) focused on stability and memory footprint, and , on a standard cpu, it tops at around 300MB/s per backup job. Your benchmarks are very interesting, and they confirm most of it. this limit was not really an issue since, in most case the xapi was limiting around 100MB/s per disk , but it will be more a more visible limit Note that master have some fixes on the memory usage (not related to backups) That's why we have started an internal workforce focused on performance, with all the teams from the kernel to the backups, including storage, network and xapi. If I can brag a little : [image: 1779106650898-afd7b59b-a4f0-4a92-88ee-2c7ba52d18bf-image.jpeg] i9 , nvme disk , backup to a nvme disk in passthrough, xoa and vm are on the same host, so it's quite far from real world data, but it shows where the limit is

@dinhngtu it still wouldn't boot with the correct template. That said, I have been testing other VMs which were created on xcp-ng, with the correct templates, and get better performance out of them when they are migrated to a hyper-v host. I believe that this is related to SMAPIv1 limitations - which are known and being actively worked on with SMAPIv3. Unfortunately for us at this point, xcp-ng may not fit our business - hopefully one day soon it will be, but until then, it is what it is.

I'm taking another look and it looks like the tools may be under another name... xen-tools I may be going blind at this point if I've missed it this many times.

@kulmacet Can you recreate this with other VMs, or just this specific oracle linux VM? I would spin up a new debian VM for example and shut this problematic VM off, and see if the issue happens with this VM as well. Outside of that, it's really looking like a hardware issue. Also, double check the ILO and BIOS firmware are at the latest. I can almost guarantee it shipped with ancient versions, and many issues like this have been patched relatively recently.