Workload-aware Page Management

[摘要] Reservation of free memory in operating systems presents a trade-off between prevention of memory shortage and providing applications a larger cache. Current operating systems use static memory thresholds to control memory subsystem behavior. For example, the free page wake up threshold determines when the page daemon will start page reclamation. Static thresholds make the memory subsystem rigid to the ever-changing system workload. With various system workloads, static memory thresholds sometimes lead to wasted memory, other times lead to a shortage of memory, harming the performance of user applications. This thesis introduces a dynamic thresholding method to balance the free memory reservation trade-off. Our system consists of two components, the allocation predictor, and ladder thresholds. The allocation predictor allows the system to set the target of free memory according to application demand. Ladder thresholds prevent incorrect predictions from leading to severe memory shortage. Our method is especially beneficial running at ;;critical points;;. For instance, we observe 14% performance boost with a database application when the database can almost fit into physical memory. Such gain can only be beneficial at those ;;critical points;;, but the performance of our method should not be worse than with the static threshold method.

[发布机构] Rice University

[效力级别] management