There were some benchmarks of DragonFly 3.0 some time ago on Phoronix. (You may recall it being mentioned here previously.) The disk numbers always seemed weird to me, so I repeated that part of the test, and here’s my writeup.
The ~10% loss in straight write performance in this case comes with GREATLY improved read/write performance and fairness. The change that likely caused this 10% loss was made knowing this would happen. While this test looks slightly worse, tests like Blogbench look far and away better. Currently we only allocate very few tags to writes so that writes do not starve reads under read/write workloads. If someone wanted to work on this perhaps we could dynamically allocate more tags in the absence of reads.
DragonFly BSD Digest