AdaptSize
AdaptSize is a cache-admission policy that adapts an exponential size bias using a model of recent object requests to improve memory-cache request hit ratio.
On a miss, an object of size s is admitted with probability exp(-s/c); s and c use the same units. A concurrent LRU variant handles eviction. The parameter is tuned from object sizes and request counts, not from a size percentile alone. [1]
The tuning model couples per-object residence probabilities through a shared pushdown rate and a byte-capacity constraint. Admission probability is not residence probability or a hit ratio. The objective counts requests equally, whereas a byte hit ratio weights payload size. [1]
In the pinned prototype, integer division inside exp changes the admission curve under LP64 types. With s = 4 bytes and c = 8 bytes, the extracted expression produces 1 instead of about 0.606531. Casting before division restores the intended value; casting afterward does not. [2] [3] [4]
Relations
Limitations
- A historical CDN memory-cache objective does not establish a universal benefit for flash caches or a guarantee of fewer backend bytes. [1]
Citations
- Berger, Sitaraman and Harchol-Balter. AdaptSize. NSDI 2017, pp. 483-498, sections 4-6.
- AdaptSize authors. Admission module, pinned commit 6deab4b303ed26c2e75d5efff6fea4e575a1485f.
- AdaptSize authors. Integer parameter declaration, pinned commit 6deab4b303ed26c2e75d5efff6fea4e575a1485f.
- Hesela. AdaptSize admission arithmetic: native extracted-expression probe, results and limitations.