Everything I've seen say using the no-op scheduler for very low latency devices is the way to go - particularly for nvme and optane devices.. interesting that it doesn't seem to be mentioned here, or in the linked paper?
Of course there also a lot of write workloads that'd still never see merges, making a scheduler just unnecessary cost.
This isn't helped by none of the current Linux schedulers really working well for local low latency storage. mq-deadline has completely antiquated heuristics and default constants, that clearly come from the spinning disk age. Bfq is too complicated (but works surprisingly well for higher latency / throttled storage).
Nor helped by the fact that nearly every storage benchmark ensures that no merging ever will occur outside of "plugging" (basically sets of IOs submitted together). No wonder they never see any benefit from schedulers :(