GL-Cache
GL-Cache is a learned cache design that predicts the usefulness of object groups and combines group selection with merge-based eviction.
Objects are grouped by write time; a learned model ranks groups, while a lightweight age-and-size rule selects survivors during merging. Group-level learning therefore does not remove object-level decisions. [1]
The utility target sums 1/(object size times next-request delay). Future requests supply training labels; eviction uses predictions, not advance knowledge. Unrequested objects can leave online labels below their complete-future values. Ghost entries extend observation after eviction. [1]
Only the first merge group is selected by utility; the paper chooses companions close in write time. Selecting the globally lowest-ranked groups is a different policy. [1]
The pinned C micro-implementation calculates training contributions using request-counter distance plus one, scaled by 1,000,000 and divided by object size. Its feature preparation separately uses real-time group age. These clocks and the label scale must not be silently interchanged when reproducing the algorithm. [2]
The author repository separates a metadata-only micro-implementation from a Rust prototype based on Pelikan Segcache. They are different evaluation surfaces, not interchangeable measurements of a deployed service. [3]
Relations
- related
- Least Hit Density, CACHEUS, Ghost cache entry, Segcache
Limitations
- The reported 228-fold throughput comparison with LRB is for the micro-implementation at small cache sizes. It excludes networking and value storage, so it is not an end-to-end service speedup. [1]
Citations
- Yang, Mao, Yue and Rashmi. GL-Cache: Group-level learning for efficient and high-performance caching. FAST 2023, pp. 115-133, sections 3.3-3.6, 4.1 and 4.4.
- GL-Cache author micro-implementation, dataPrep.c: prepare_one_row and update_train_y; commit fbb824091209f390f7c921aadee1fbf0d35209c2.
- GL-Cache author repository: micro-implementation versus prototype; commit fbb824091209f390f7c921aadee1fbf0d35209c2.
Sources & caveats
- Primary-paper and pinned-source review. No trace replay, native benchmark, model training or deployment-performance replication was performed.
- The record distinguishes the paper's design from the released micro-implementation; code-specific clocks are not attributed to every implementation.