Average eviction time (AET)
Average eviction time is the mean interval from a cached block's last access to its eviction, used by a kinetic model to estimate LRU miss-ratio curves.
AET models a fully associative LRU cache using the reuse-time tail P(t). It finds a cutoff T whose tail integral equals cache capacity, then estimates the miss ratio as P(T). Time counts references, not seconds or storage latency. [1]
Eviction time runs from a block's last access to its eviction, not from insertion. In two-slot LRU on ABACD, A enters at index 0, is reused at 2 and leaves at 4: eviction time is 2, residence time is 4. Residents still present at the trace end have no observed completed eviction. [1] [3]
Cold accesses contribute infinite reuse times to the histogram. This keeps their probability in the tail. Phase sampling retains outstanding monitors across phase boundaries; resetting them discards cross-phase reuse. [1]
Approximation matters: AACBAB and ABAABC each have three cold accesses and finite intervals 1, 2 and 3. A declared step-tail integration gives T = 9/4 at capacity 2 and predicts 4/6 misses for both. Exact empty-cache LRU instead gives 4/6 and 3/6. The shared histogram cannot distinguish the traces. [3]
This calculation is an illustrative interpretation of the paper's equations, not its sampler or benchmark implementation. Sampling error and histogram-to-LRU approximation are different issues; full-trace inputs do not make the latter universally exact. [1] [4]
Relations
- related
- Reuse time, Miss-ratio curve, StatStack
Citations
- Hu et al. Kinetic Modeling of Data Eviction in Cache. USENIX ATC 2016, pp. 351-364; sections 2.1-2.3 and 3.2-3.3.
- Hesela. AET step-tail arithmetic and completed-eviction ledgers; constructed traces.
- Hesela. Bounded evaluator with exact rational cutoffs and explicit interpolation.
Further reading
- USENIX. Kinetic Modeling of Data Eviction in Cache: authors and proceedings metadata.
- Hesela. Independent tail-sum and stack-distance checks for the teaching model.
Sources & caveats
- Primary paper sections 1-7 reviewed. No author implementation, sampler, hardware result or original workload was reproduced.
- The 2018 journal extension is separate: Fast miss ratio curve modeling for storage cache, DOI 10.1145/3185751; this record describes the 2016 paper.
- The two histogram-equivalent traces reuse existing Hesela evidence. New arithmetic declares its interpolation and preserves the all-request denominator.