Swapping to conventional SSDs suffers from opaque flash management: the device’s own garbage collection runs uncoordinated with the OS, causing write amplification, throughput cliffs, and poor isolation between tenants. ZNSwap redesigns the Linux swap subsystem around Zoned Namespace (ZNS) SSDs, moving garbage collection into the host and co-designing it with the kernel swap logic so the two share usage statistics and correctly account for GC bandwidth. The result is stable throughput across access patterns, up to 10× lower tail latency, and 5× higher throughput for an in-memory key-value store.

Published at USENIX ATC 2022 and in ACM Transactions on Storage 2023.