Simulation parameters.

We used four different target compression sizes with values equal to 10%, 23%, 37%, and 50% of the simulated memory size. Thus, during persistent phases of program behavior (i.e., when the system has enough time to adapt) either none, or 10%, or 23%, or 37%, or 50% of our memory pages are holding compressed data. Limiting the number of target compression sizes to four guarantees that our cost/benefit analysis incurs a low overhead. The decay factor used is such that the M-th most recent event (with M being the size of memory) has a weight equal to 20% of the most recent event. Our results were not particularly sensitive to the exact value of the decay factor.