Explore a 27-minute conference talk from USENIX FAST '20 that delves into FPGA-accelerated compactions for LSM-based key-value stores. Learn how researchers from Alibaba Group and Zhejiang University address performance challenges in Log-Structured Merge Tree (LSM-tree) key-value stores by offloading compactions to FPGAs. Discover the impact of slow compactions on system performance, the bottleneck caused by CPU-bound operations when merging short key-value pairs, and the proposed solution to accelerate compactions and reduce CPU contention. Gain insights into the implementation of this approach in X-Engine, a modern LSM-tree storage engine, and understand the significant improvements in compaction speed, system throughput, and energy efficiency compared to CPU-only baselines.
Overview
Syllabus
FAST '20 - FPGA-Accelerated Compactions for LSM based Key-Value Store
Taught by
USENIX