Far into the C3 project, someone asked us how we knew we could pay all the employees in X time, given that now it took 8X or something. 

We ignored them for a long time, saying that after all we had a 10 CPU machine and were only using one of them. Finally someone of high enough rank asked loudly enough, saying they didn't believe that the ten CPUs would make it any faster.

We took a morning, split the population into 8 groups, started 8 tasks, ran them in parallel. It ran in X. We reported the results the next day, including the ''fact'' that we ran the pay in 1/8 the time and got the same answer, and the 
''estimate'' that it would take two weeks to actually set up the system to do it automatically.

The problem went away for several months, until actual run time was painful. Then the customers scheduled the task and we did it.

The morning's work answered the question. The rest was just refinement. PerformanceSpike.  --RonJeffries