SieveStore
SieveStore is a shared block-cache design that filters admissions to reduce cache-fill writes while retaining frequently accessed blocks.
SieveStore-D selects blocks from access logs at epoch boundaries. SieveStore-C admits blocks after repeated misses, using an imprecise first-stage counter table followed by precise per-block tracking. The latter prevents hash collisions alone from qualifying low-reuse blocks. [1]
A read miss can cause a write when its fetched block enters an SSD cache. Choosing an eviction victim does not eliminate that admission write. Cache-fill writes and internal NAND copying are separate accounting layers. [1] [3]
The 2010 SieveStore design is distinct from the 2024 SIEVE eviction algorithm. [1] [2]
Relations
- related
- Cache admission, Cache pollution, SIEVE cache eviction, Write amplification, Least recently used replacement
Limitations
- The evaluation used a week of block traces from 13 servers. Its minute-level drive-demand model ignores queueing and does not establish request tail latency. [1]
Citations
- Pritchett and Thottethodi. SieveStore. ISCA 2010, pp. 163-174, sections 3-5 and Table 2. DOI: 10.1145/1815961.1815982.
- Zhang et al. SIEVE is Simpler than LRU. NSDI 2024, pp. 1229-1246.
- Hesela: SieveStore allocation-write accounting and the total-versus-increment denominator.
Sources & caveats
- Full ISCA 2010 paper reviewed, including the accounting table and evaluation assumptions. The linked code reconstructs Table 2 arithmetic, not the trace-driven simulator or an SSD benchmark.