Online Working Set Change Detection with Constant Complexity
L J Gokul Vasan, Éder F. Zulian, Christian Weis, Matthias Jung, Norbert Wehn · 2021
In advanced scalable multiprocessor systems with multiple applications running concurrently, the optimal utilization of memory resources plays an essential role in achieving high system performance and energy efficiency. Locality of reference is one of the basis of memory management techniques. In operating systems, the memory management benefits from temporal and spatial locality when deciding which page shall be moved down the memory hierarchy to make room for a new page. However, current algorithms are oblivious to the moments when individual tasks transition to a new working set, which leads to suboptimal decisions. Furthermore, the eviction of idle pages from main memory is postponed to the last minute, which leads to system unresponsiveness and thrashing. In this context, memory locality reasoning based on the phase transition model is a central aspect to consider. We perceive that incorporating such a model in memory management algorithms provides the operating system with ways to trigger proactive page reclamation or compaction, to collect memory related statistics, and to trigger other crucial decision-making.