Storage Gardening: Using a Virtualization Layer for Efficient Defragmentation in the WAFL
summary
WAFL leverages a V→P indirection and a fat container file to move physical blocks (including snapshot leaves) without rewriting file-level pointers, implementing segment cleaning, just-in-time cleaning, and sub-block compaction to reclaim free space and reduce read/write IOPS. Evaluation shows HDD aggregates need ~60 days of workload to measure benefits and stabilize (about 35 days with cleaning enabled), SSDs fragment in hours–days, and all‑flash systems sometimes see worse tail latency if defragmentation runs indiscriminately. Techniques are implemented in production and reduce fragmentation when paced to workload.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:58What is fragmentation
- 2:24Background
- 3:56Layering
- 6:02Number Spaces
- 15:19Evaluation
- 18:30Subblock compaction
- 20:37Rhe compaction
- 21:41Fourth form fragmentation
related talks
48mOptimal SSD Management with Predictions
34mSustaining Git Performance Under Heavy Workloads
18mOxbow: A Coordinated Architecture for Multi-Component File Systems
48mDesign of data-intensive applications
42mAdventures in Efficiency and Performance
18mHarvesting Idle Memory for Application-managed Soft State with Midas
16mMemory-Efficient Fixpoint Computation
13mEfficient data aggregation with Fenwick trees