Cost for Finding Peak Rate

Figure 7 shows the choice of load factors for finding the peak rate for a sample with $4$ disks and $32$ nfsds using the policies outlined in Section 4. Each point on the curve represents a single trial for some load factor. More points indicate higher number of trials at that load factor. For brevity, we show the results only for DB_TP. Other workloads show similar behavior.

For all policies, the controller conducts more trials at load factors near $1$ than at other load factors to find the peak rate with the target accuracy and confidence. All policies without seeding start at a low load factor and take longer to reach a load factor of $1$ as compared to policies with seeding. All policies with seeding start at a load factor close to $1$, since they use the peak rate of a previous sample with $4$ disks and $16$ nfsds as the seed load.

Linear takes a significantly longer time because it uses a fixed increment by which to increase the test load. However, Binsearch jumps to the peak rate region in logarithmic number of steps. The Model policy is the quickest to jump near the load factor of $1$, but incurs most of its cost there. This happens because the model learned is sufficiently accurate for guiding the search near the peak rate, but not accurate enough to search the peak rate quickly.

Figure 7: Time spent at each load factor for finding the peak rate for different policies for DB_TP with $4$ disks and $32$ nfsds. Seeded policies were seeded with the peak rate for $4$ disks and $16$ nfsds. The result is representative of other samples and workloads. All policies except linear quickly converge to the load factor of $1$ and conduct more trials there to achieve the target accuracy and confidence.