Explore the Zoned Namespace (ZNS) interface for flash-based SSDs in this USENIX ATC '21 conference talk. Discover how ZNS offers a solution to the rising block interface tax by exposing flash erase block boundaries and write-ordering rules. Learn about the impact on SSD hardware/firmware and host software, and understand how storage software can be specialized for ZNS semantics. Examine the modifications required to support ZNS SSDs and see how adapted versions of f2fs and RocksDB leverage ZNS to achieve higher throughput and lower tail latency compared to traditional block-interface SSDs. Gain insights into the significant efficiency benefits, including 2–4× lower 99.9th-percentile random-read latency and 2× higher write throughput for zone-specialized RocksDB on ZNS SSDs.
ZNS - Avoiding the Block Interface Tax for Flash-based SSDs
Overview
Syllabus
Intro
Zoned Namespace SSDs Getting rid of the block interface tax
Evolving towards ZNS SSDs
Adoption Three ways to adopt ZNS SSDs
Enabling the Linux Ecosystem Adding support for ZNS SSDs
ZenFS Architecture A new storage backend for RocksDB
Evaluation Apples-to-apples comparison
Raw 1/0 Characteristics Improving Write Throughput & Read Latency
RocksDB: Writes Double the throughput over 28% OP SSDs
RocksDB: Reads and Writes Improving Writes and Tail Latencies
Summary
Taught by
USENIX
