The current WiredTiger eviction worker dynamic allocation algorithm is quite passive, which means that short run benchmarks can show poorer than ideal performance. The YCSB load phase via MongoDB is an example, if a load of approximately 1 minute is configured, and MongoDB is running for at least 30 seconds prior to the benchmark starting only the eviction server is ever run.

We should revisit after WT-3123 is merged - it may be practical to more aggressively start and stop threads, if that change makes start/stop more light weight.