XFS, 4 NVMe striped devices, stripe size 64k, hyperthreading disabled I/O througput on XFS. We see that there is quite an overhead, it doesn't reach the levels in fio-striped.txt. QAT has lower througput than AES-NI, but it also has very low CPU consumption when reading - this is the only dm-crypt test where QAT provides some benefit! fio --ioengine=psync --iodepth=1 --rw=randread --direct=0 --end_fsync=1 --bs=64k --numjobs=56 --time_based --runtime=10 --group_reporting --name=job --filename=/mnt/test/file fio --ioengine=psync --iodepth=1 --rw=randwrite --direct=0 --end_fsync=1 --bs=64k --numjobs=56 --time_based --runtime=10 --group_reporting --name=job --filename=/mnt/test/file raw device: READ: bw=3672MiB/s busy: 45.963518% idle: 53.343391% irq: 0.436703% WRITE: bw=2966MiB/s busy: 57.770084% idle: 28.695388% irq: 0.983228% dm-crypt, 4096, aes-ni READ: bw=2956MiB/s busy: 60.790450% idle: 38.079135% irq: 0.660701% WRITE: bw=3156MiB/s busy: 59.604016% idle: 35.336383% irq: 0.815776% dm-crypt, 4096, qat READ: bw=1678MiB/s busy: 4.079595% idle: 92.481173% irq: 1.898289% WRITE: bw=1805MiB/s busy: 29.078186% idle: 62.963860% irq: 2.034750%