In an extended investigation (started via a skunk project) into using "abbreviation tables" for faster searching, a microbenchmark was developed that mimics the key characteristics (key length, distribution) and workload mix (search vs insert, etc) that we see typically in mongodb applications, and also specifically in our high value workloads.

This new benchmark enabled the discovery of some small changes to btree searching code for row store. Taken together we see ~4-5% throughput gains with the mixed workloads in the benchmark, and more like ~7-10% gains for benchmark workloads that are more read-focussed.  The usual caveats apply, as the benchmark is very cursor centric - keeping a cursor open throughout.  Fold in the cursor opens, transactions, timestamp, multiple btrees, mongodb layering, etc. and we'll likely see a dilution of these numbers in practice.

This ticket is to add the new benchmark, and the known wins, and provide supporting evidence with evergreen runs.