hesela.devBlock Padding LRU (BPLRU)
Block Padding LRU (BPLRU) is a research SSD write-buffer policy combining block-level recency, victim padding and sequential-write compensation.
Buffered sectors share a block-level LRU position. Eviction can read missing clean pages and write a complete block sequentially, enabling a switch merge in the studied log-block FTL. Fully sequential blocks move toward eviction. Reordering can help even with zero sector-cache hits. [1]
The simulation omitted RAM-copy cost; the 1 GB MLC prototype's MP3-copy case was slower than no buffer. Flushes constrain buffering benefits. These are important counterexamples to equating buffer size or cache hits with performance. [1]
A two-level index locates a resident sector with two index references. Block headers and sector nodes use a shared free-node pool, so metadata consumes buffer memory. [1]
In the uniform-random W4 simulation, BPLRU erased fewer blocks yet ran more slowly than FAB, which selected victims by buffered-sector count. Lower erase count did not imply higher throughput. [1]
Relations
- related
- Write combining, Flash translation layer (FTL), SSD DRAM cache, Least recently used replacement, Storage cache flush, Write amplification
Citations
Sources & caveats
- Historical simulation/prototype evidence, not a modern-SSD performance guarantee.
- No implementation or benchmark was reproduced by Hesela.