Sue's measurements show that adding extra LSM merge threads slow down applications.

If the disk is the bottleneck, this may be reasonable: multiple merge threads will do more I/O that a single merge thread.

Check if disk is the bottleneck, or whether there is some other lock contention involved.

Think about whether we can just have one flavour of LSM worker thread that does switching, flushing, Bloom filter creation or merging as needed.