Hi, Currently, I'm testing Ceph v17.2.7 with NVMe. When mapping an rbd image to physical compute host, "fio bs=4k iodepth=128 randwrite" give 150k IOPS. I have a VM that located within the compute host, and the fio give ~40k IOPS and with 50 %iowait. I know there is a bottleneck, I'm not sure where it is though I have tried to enabled iothread in virsh but nothing change. Anyone have any ideal? Thanks in advance.