Abstract
The classical working set (WS) model limits thrashing by keeping in memory all pages referenced within a fixed window delta. However, a constant delta cannot adapt to changing locality across execution phases, leading to either inefficient memory utilization or increased page faults. This paper treats delta as a controlled variable rather than a configuration constant. We define a feedback controller that samples the fault rate over fixed reference epochs and enlarges or reduces delta by a bounded step when the rate crosses an upper or a lower threshold. A dead band between the thresholds, a bounded adjustment step and a cooldown interval together bound the adaptation rate and exclude sustained oscillation. The controller needs no offline profiling and no per-workload parameter search, and it leaves the eviction rule itself unchanged, so setting the step to zero recovers the classical policy exactly and any measured difference isolates the effect of adaptation alone. We evaluate the policy in a trace-driven simulator against least recently used (LRU) and first-in, first-out (FIFO), not recently used (NRU), and the static WS policy on phase-changing, cyclic and random-uniform reference streams, reporting fault counts with confidence intervals. We also state the regimes in which adaptation is expected to give no benefit.