iotune may use a too-low iodepth

[travisdowns]

iotune uses nr_requests to select the iodepth to use for its IOPS-related benchmarks: it sums the nr_requests value from the involved disks (including the underlying devices if md is being used). However, nr_requests is not an upper bound on the amount of requests that can be in flight in a disk: see here for example:

... Note that nr_requests only applies to the I/O operations within the I/O scheduler and not to I/O operations already dispatched to the underlying device. Therefore, the maximum outstanding limit of I/O operations against a device is (nr_requests*2)+(queue_depth) where queue_depth is /sys/block/sdN/device/queue_depth, sometimes also referred to as the LUN queue depth. You can see this total outstanding number of I/O operations in, for example, the output of iostat in the avgqu-sz column.

This is actually noticeable when testing the same devices with fio: performance may keep increasing even as iodepth rises above the nr_requests value. Perhaps we should use the sum of the block device queue size + nr_requests. Or make it configurable.

[mykaul]

Ref https://github.com/scylladb/seastar/issues/1297

[mykaul]

CC @xemul