Venue | Category |
---|---|
FAST'19 | PM |
SLM-DB: Single Level Merge Key-Value Store with Persistent Memory1. SummaryMotivation of this paperSLD-DBImplementation and Evaluation2. Strength (Contributions of the paper)3. Weakness (Limitations of the paper)4. Some Insights (Future work)
LSM-Tree
LSM-Tree limitations
Large overhead to locate needed data
High disk write amplification
Persistent Memory + Disk
Architecture
Persistent MemTable
Persistent B+-tree Index (FAST-FAIR B+-tree)
Selective compaction
why compaction?
Selectively pick SSTable files
Make those files as compaction candidates
Merge together most overlapping compaction candidates
Selection schemes for compaction candidates:
Compaction triggered when there are too many compaction candidate files
Evaluation
PM simulator: PM write latency 500ns (5x of DRAM write latency)
Intel’s PMDK used to control PM pool
db_bench and YCSB
microbenchmark
PM sensitivity
A new design of KV-stores with persistent memory
Low write amplification and near-optimal read amplification
Level-DB
Memory:
MemTable, Immutable MemTable
In-memory skiplist to buffer updates, mark immutable when becoming full
sequential write to the disk
Disk:
LSM-tree optimizations