DiLOS: Adding Performance to Paging-based Memory Disaggregation

Memory disaggregation places computing and memory in physically separate nodes and achieves improved memory utilization in datacenters. Kernel-based approaches for memory disaggregation offer transparent virtual memory by using paging schemes but suffer from expensive page fault handling. As an alternative, library-based approaches incorporate application semantics to memory disaggregation and can even eliminate page fault handling on its data path. However, its lack of compatibility harms generality and obstruct wide adoption. This paper revisits the paging-based approaches and challenges their performance. We posit that the page fault overhead is not a fundamental limitation. We propose DiLOS, a new memory disaggregating unikernel, that delivers both performance and generality. The key insight of DiLOS to overcome performance drawbacks while maintaining generality lies in the design of a fast, lightweight page fault handler on top of the unikernel's simple execution model. Since each unikernel serves a single application, it also opens room for extra optimization via app-aware prefetching. Di-LOS outperforms a recent library-based system (AIFM) by 1.52x and 1.31x when the cache size is 12.5% and 100% of the total working set, respectively. Compared to the state-of-the-art paging-based system (Fastswap), DiLOS with a general-purpose prefetcher achieves up to 2.2x higher performance in real-world workload. An app-aware prefetcher further improves the throughput of Redis in-memory database up to 27%.