Pay migration tax to homeland: Anchor-based Scalable Reference Counting for Multicores

Abstract

The operating system community has been combating scalability bottlenecks for the past decade with victories for all the then-new multicore hardware. File systems, however, are in the midst of turmoil yet. One of the culprits behind performance degradation is reference counting widely used for managing data and metadata, and scalability is badly impacted under load with little or no logical contention, where the capability is desperately needed. To address this, we propose PAYGO, a reference counting technique that combines per-core hash of local reference counters with an anchor counter. PAYGO imposes the restriction that decrement must be performed on the original local counter where the act of increment has occurred so that reclaiming zero-valued local counters can be done immediately. To this end, we enforce migrated processes running on different cores to update the anchor counter associated with the original local counter. We implemented PAYGO in the Linux page cache, and so our implementation is transparent to the file system. Experimental evaluation with underlying file systems (i.e., ext4, F2FS, btrfs, and XFS) demonstrated that PAYGO scales file systems better than other state-of-the-art techniques.