hesela.devLeCaR
LeCaR is a cache replacement policy that learns a probability distribution over LRU and LFU eviction choices from remembered eviction mistakes.
An eviction expert is sampled using its current weight, rather than always choosing the larger weight. A miss on a remembered victim penalizes the responsible expert; the original design discounts this penalty with time spent outside the cache and uses a fixed learning rate. [1]
Both experts describe the same resident cache, not two independently running caches. In the pinned implementation, agreement on a victim bypasses attribution history; an entry forgotten from bounded history also cannot trigger a later weight penalty. Absence of a penalty therefore does not certify a good eviction. [3]
CACHEUS revisits the expert pair and adapts the learning rate to address workload-dependent weaknesses. [2]
Relations
- related
- CACHEUS, Least recently used replacement, Ghost cache entry, Adaptive Replacement Cache, Cache pollution, Least frequently used replacement
Citations
- Vietri et al. Driving Cache Replacement with ML-based LeCaR. HotStorage 2018, sections 3-4.
- Rodriguez et al. Learning Cache Replacement with CACHEUS. FAST 2021, pp. 341-354, sections 3-6.
- LeCaR author implementation: evict, addToHistory, miss and adjustWeights; commit 1eec63ce166502be33ddd1f35bc041ed73a24f4d.
Further reading
Sources & caveats
- Primary-paper and pinned-source review. No author simulator execution or reproduction of workload hit rates.
- The consensus-eviction detail is scoped to the linked implementation, not asserted as a line in the original LeCaR pseudocode.