Swap Fairness for Thrashing Mitigation

Overview & Synopsis

The swap mechanism allows operating systems to manage more memory than the available RAM space, by temporarily storing unused memory pages on disk. However, disk transfers are way slower than normal RAM operations, and under memory pressure, the system may spend more time in retrieving and storing swapped pages than performing actual computation: this state is called memory thrashing. To reduce thrashing, several ideas were brought up to optimize page replacement algorithms and system-wide load. For instance, Linux currently implements the swap-token, a mechanism designed to immunize the memory pages of the heaviest process against swapping. Such a mechanism eliminates early thrashing peaks and improves general system performance if the process is to finish quickly. The swap-token may however be counterproductive when it is tricked into advantaging malicious or long-standing processes. This is particularily true in the context of shared hosting or virtualization, where multiple users run uncoordinated and selfish workloads. In this paper, we present an accounting layer that forces swap fairness among processes competing for main memory. It ensures that a process cannot perform longer swap operations than others, and delays the swap operations of processes abusing the swapping mechanism. With such a layer, we are able to significantly reduce the dispersion of execution times under memory pressure, and generally improve the performance of legit, memory-heavy processes running concurrently with abusive ones.