hesela.devCACHEUS
CACHEUS is an adaptive cache replacement framework that learns how to choose between two eviction experts while adjusting its learning rate.
The C3 variant combines scan-resistant SR-LRU with churn-resistant CR-LFU. SR-LRU separates reused entries from a filtering partition; CR-LFU breaks lowest-frequency ties by evicting the most recently used candidate. [1]
Unlike LeCaR, CACHEUS removes the time discount on remembered mistakes and adapts learning rate over request windows. Expert weights are selection probabilities, not hit probabilities or capacity partitions. If both experts nominate the same victim, no attribution-history entry is added. [1] The pinned implementation retains this consensus branch. [2]
The evaluated expert combinations have workload-dependent outcomes; online learning does not establish universal dominance over other policies. [1]
Relations
- related
- LeCaR, Least recently used replacement, Ghost cache entry, Low inter-reference recency set replacement, Cache pollution, Least frequently used replacement
Citations
- Rodriguez et al. Learning Cache Replacement with CACHEUS. FAST 2021, pp. 341-354, sections 3-6.
- CACHEUS author implementation: eviction attribution and adaptive learning rate; commit 1eec63ce166502be33ddd1f35bc041ed73a24f4d.
Further reading
Sources & caveats
- FAST 2021 mechanism and evaluation review, with source inspection at a pinned commit. This record does not reproduce the simulation results.