I saw a similar effect with really fast disks. If you make the kernel faster at passing requests down to the disk, a simple benchmark with one request at a time

I saw a similar effect with really fast disks. If you make the kernel faster at passing requests down to the disk, a simple benchmark with one request at a time will be faster. However, with many requests at the same time and less time spent processing then, you now have more time to poll the disk for completed requests. Each time the kernel polls the disk, it will typically see fewer completed request than before your optimizations, and overall this can actually result in decreased throughput.