Utilization, capacity, and waiting time
Relate demand to service capacity and understand why waiting rises before a system reaches 100% average load.
For identical parallel servers, a rough offered-load calculation compares the arrival rate with total mean service capacity. If one server completes μ jobs per hour on average and there are c servers, total mean capacity is cμ.
offered load ρ = λ / (cμ)
Arrivals do not come at perfectly even intervals and jobs do not all take the same time. Bursts and long jobs create temporary backlogs, so waiting can become large as spare capacity shrinks.
That is why an average load below 100% does not mean nobody waits.
Two of three servers busy right now means current occupancy is 66.7%. Utilization is normally a time-based measure over an interval. Do not label a single instant as if it were a long-run utilization estimate.
When diagnosing a queue, look at waiting time, queue length, throughput, and time-based utilization together.
Check every stage in the path. If an upstream process can only supply 9 jobs per hour, increasing a downstream station from 11 to 15 jobs per hour will not increase total throughput.