A typical workflow for wtperf is: (1) it fills in the database (2) runs one or multiple workloads on it.

The issue here is that after after filling in the database, a part of it remains in the OS cache leading to unfair results because part of supposedly "cold" data in fact doesn't get read from disk.

This also can impact the test results if the workload uses an existing database.

wtperf must clear the OS cache before doing any performance measurement runs.

• Cleaning cache on Linux (clear only PAGE cache):

  sudo bash -c 'sync; echo 1 > /proc/sys/vm/drop_caches'
  

  (link1 link2)

• Cleaning cache on MacOS:

  sync ; purge
  

• Cleaning cache on Windows:
  ?