Looking at individual applications, however, we see a couple of outliers. In cg, both Ipek+PwDn/Up and MORSEP derive noticeable benefit from having PwDn/PwUp available, even if MORSE-P still edges out Ipek+PwDn/Up significantly (Figure 3). cg strongly favors an open page policy, as evidenced by the large expectation of a page hit for FR-FCFS (which is open-page). Ipek destroys such potential by being forced by design to close pages prematurely over doing nothing (Figure 4). On the other hand, swim seldom hits on open pages for FR-FCFS, and actually performs best with Ipek, where the NoOp restriction is unavoidable
and it results in a high number of preemptive precharges. Relaxing this constraint actually hurts Ipek+PwDn/Up and MORSE-P virtually equally with respect to Ipek, although they still outperform full-open-page FR-FCFS significantly. In other words, while these configurations do learn to preemptively close pages and derive speedups as a result, in the case of swim they fall short of the level of aggression with which they could apply it, as (accidentally) evidenced by Ipek. (Note that swim is not part of our training set. While the obvious temptation is to include it, this would amount to overfitting.)