MatrixKV - Reducing Write Stalls and Write Amplification in LSM-tree Based KV Stores

Explore a conference talk from USENIX ATC '20 that presents MatrixKV, an innovative approach to improving performance in LSM-tree based key-value stores. Delve into the challenges of write stalls and write amplification in popular key-value stores, and discover how MatrixKV leverages non-volatile memory (NVM) to address these issues. Learn about the four novel techniques introduced: the matrix container for managing L0 level in NVM, column compaction for fine-grained L0 to L1 compaction, increased level width to reduce LSM-tree depth, and cross-row hint search for maintaining read performance. Gain insights into the implementation and evaluation of MatrixKV, which demonstrates significant improvements in latency and random write throughput compared to RocksDB and NoveLSM. Understand the potential impact of this research on the performance and efficiency of key-value stores in multi-tier DRAM-NVM-SSD storage systems.

Intro
Outline
Challenge 1: Write stall
Root cause of write stall: LO-L1 compaction
Root cause of increased write amplification
Motivation
Overall Architecture
Matrix Container
Row Table
Fine grained column compaction
Reducing LSM-tree depth
Cross-Row hint search
Evaluation Setup
Random Write Throughput
Write stalls
Tail Latency
Fine granularity column compaction