In WT-3725 I added various statistics around timestamp usage including tracking around three transaction subsystem rwlocks: txn_global->rwlock, commit_timestamp_rwlock, read_timestamp_rwlock. Although the code for lock tracking now uses the rdtsc code instead of wt_epoch there is still a consistent 2-3% degradation in the insert workload in SERVER-31694 when timestamp locks are tracked versus when they're not tracked.

Reproduce the regression, determine if there is an obvious way to avoid it, and if not, check whether the statistics we have are sufficient.