I'd say the MongoDB folks made a pretty safe choice here, since WiredTiger offers both LSM and Btree engines. I've recently completed some more benchmarks on all of the above-tested engines, and will be publishing the report soon (as soon as it's done being written...). Testing on-disk, with DBs 5x larger than RAM, at data sizes of 24, 96, 384, 768, 2000, and 4000 bytes per record. The important result to note is that LSM's high write speed advantage only exists for small record sizes, and for short durations. On long duration tests the compaction overhead stops them in their tracks, and at data sizes above 2000 bytes per record, LSM write amplification is too expensive (regardless of test duration). I.e., above 2000 byte records, Btrees are always faster than LSMs.

This result is particularly relevant for MongoDB, since a document store tends to have large records containing multiple fields, as opposed to individual values being stored as separate small records.

And in cases where that applies (large records), nothing comes anywhere close to LMDB's performance.

(A preview of the on-disk test results I mentioned was presented at BuildStuff.LT last week http://symas.com/mdb/20141120-BuildStuff-Lightning.pdf pages 103-on)