Scheduler-Aware Virtual Memory Management

Virtual memory management has come to play an increasingly adversarial role in the scheduling of processes. While it is the job of the CPU scheduler to determine the order and length of execution of processes, its attempts to meet the desired scheduling constraints are thwarted by the memory manager under memory pressure. This is because scheduler policies like priority are not supported by memory managers, so that heavy paging under memory pressure can result in non-deterministic divergence from the desired CPU allocations. For instance, the memory access pattern of a low-priority process could cause a higher-priority process to lose its pages and page fault frequently, thus inducing a form of priority inversion. This effect is the result of two problems with traditional virtual memory managers. First, they are scheduler-oblivious, i.e., they are neither aware of scheduler policies, nor do they attempt to cooperate with the scheduler. Second, their primary objective is to maximize system-wide throughput by maximizing the overall page faults. This approach prevents the meeting of scheduler goals in the presence of memory pressure. In this work, we describe SAVMM – the first scheduler-aware virtual memory manager that we are aware of. SAVMM works by formalizing the notion of scheduler-awareness as a utility maximization problem, where the utility of a process is based on its memory requirement as well as its scheduler-defined importance.