This is a follow-on to
SERVER-20185. See that ticket for test details, but in brief each connection is doing 10 inserts per second, and throughput is measured as the number of connections is slowly ramped up. At high connection counts negative scaling is observed. The same test at low connection count (100-200 connections) but no delay in the loop (i.e. each connection doing inserts flat out) yielded ~260k inserts/s, so the expected 150k inserts/s at 15k connections should be within the raw capacity of the system.