> Traditionally the reason to use threads in small-requests i/o is to get multiple outstanding requests in flight in the hardware io stack which some hardware c

> Traditionally the reason to use threads in small-requests i/o is to get multiple outstanding requests in flight in the hardware io stack which some hardware can use to get you more IOPS throughput I would assume that is the case here too. The threads are likely hiding the latency of the IOPS. I can't see any CPU load statistics, so it is hard to say, if it is indeed CPU bound or not. A quick run with time shows user+sys being only slightly higher than real, so with 6 threads for my 6 cpus, it uses a single cpu. (and practically no time in user-space)