ScaleSwap: A Scalable OS Swap System for All-Flash Swap Arrays

초록

This paper presents a scalable OS swap system, ScaleSwap, designed to enhance core and SSD scalability on all-flash swap arrays. Specifically, ScaleSwap first enables a one-to-one swap model where each core exclusively manages its own swap resources, enabling core-centric swap in/out operations. Second, ScaleSwap devises opportunistic inter-core swap assistance, allowing each core to delegate swap metadata access to other cores as needed. Finally, ScaleSwap adopts core-affinity page and LRU management to mitigate LRU lock contention during swap in/out operations. We implement ScaleSwap in the Linux kernel and evaluate its performance on a 128-core machine with an all-flash swap array comprising eight NVMe SSDs. Our evaluation shows that ScaleSwap achieves up to 3.4× higher throughput and up to 11.5× lower average latency than the Linux swap system. Furthermore, ScaleSwap outperforms two prior systems, TMO and ExtMEM, by up to 64% and 5×, respectively.