Given this constraint, our model identified an optimal configuration
for each workload as shown in
Figure 7b. We can see that the optimum number of
active tips is at the top of the range for the values of that
parameter. This is not surprising, because additional active tips
lower service time. The movement ranges in and
are small,
which also reduces service time. If we compare the values of the
optimal configurations to the behavior of the experimentally obtained
values for the service time in Figure 7a we can see
that our predictions point out very well where the minimum of the
service time will occur.
Note that the error is greatest for the tpcd workload, where the model
generally underestimates the service time. Specifically, the model
underestimates the seek time for tpcd because the distribution of
seeks in the real workload does not fit our assumption of a uniformly
random distribution. The tpcd workload has a significant number of
large seeks that increase the overall average. As we change the
movement range in and
, the model prediction changes more
quickly than the real workload. Thus, at higher values of
,
the simulated and predicted service times converge.