Explore a conference talk on MadFS, a novel approach to per-file virtualization for userspace persistent memory filesystems. Learn about the challenges of managing persistent memory in userspace and how MadFS addresses them through embedded metadata, copy-on-write for crash consistency, and lock-free optimistic concurrency control. Discover how this innovative filesystem achieves significant performance improvements over existing solutions, with up to 3.6x throughput compared to ext4-DAX and substantial speedups for real-world applications like YCSB on LevelDB and TPC-C on SQLite. Gain insights into the design, implementation, and evaluation of MadFS, including its simplified architecture, metadata management techniques, and performance in concurrent workload scenarios.
MadFS - Per-File Virtualization for Userspace Persistent Memory Filesystems
Overview
Syllabus
Intro
Background: Persistent Memory
Background: Kernel Filesystems for PM
Background: Userspace Filesystems for PN
MadFS: Metadata Embedded Filesystem
MadFS: Simplified Design
MadFS: Metadata Management
MadFS: Per-File Virtualization
Evaluation: Concurrent 4 KB Random Read
Evaluation: Concurrent 4 KB Random Overwrite
Evaluation: TPC-C on SQLite
Taught by
USENIX
