Tiered-memory optimizations were designed for local byte-addressable non-volatile memory that is only a few times slower than DRAM. Disaggregated memory is far slower, which undermines those well-established techniques. Through analysis of real workloads, this work shows that disaggregation calls for different design choices, and builds HotBox, a Linux disaggregated-memory management subsystem that maximizes the local-memory hit rate with low overhead and only minor kernel changes — outperforming state-of-the-art tiered-memory systems by up to 2.25× on memory-intensive benchmarks.

Published at ACM SIGPLAN ISMM 2022.