...
For 16 way parallel MPI jobs we'd typically expect a CPU core account mix of around 95% User, 5% System, 1%-5% IO Wait and 0% Idle time. When a job exhibits the pathology the primary evidence is a mix of 20-50% User, 10-20% System, 0-1% IO Waid Wait and 30-70% Idle time. The high idle time is the most curious aspect.
...