There have been a few MongoDB workloads that cause a cascading performance degradation in WiredTiger. The application is attempting to push through more workload than the system can handle, as application threads become co-opted into eviction the throughput rate reduces further.
When the throughput rate drops, the application opens more threads to the database, attempting to keep the amount of throughput stable. As the number of threads grows, WiredTiger successfully completes less and less work.
The bottle neck appears to be eviction. We should create a workload in standalone WiredTiger that replicates the negative scaling.