Hybrid XFS: Idea to Exabytes in 12 Months
summary
Claim: hybrid XFS converts XFS real-time to use an SSD for metadata and HDDs for data and rolls out to a multi‑exabyte fleet in ~12 months. Uses blocktrace/FB outrace for IO analysis and FBar automation to drain, recreate, and remount; control group testing shows 15–25% latency reduction and about a 30% I/O efficiency uplift. Implementation details: RT inherit flag enabled, extent size moved from 256K to 1M, 1TB SSD serving ~36 HDDs, small‑file cutoff ~64KB, and a metadata rescue partition for SSD failure.
AI-written summary · about
chapters (28)
- 0:00Intro
- 0:44What is your job
- 1:22Factors that make this job tough
- 2:13AIOPS
- 2:39filesystem behavior
- 3:32block trace
- 3:57page cache
- 5:59realtime subvolumes
- 7:18realtime allocator
- 9:02changes
- 10:07patches
- 10:32deployment
- 11:34blocktrace verify
- 11:50contingency plan
- 12:55proofofconcept
- 13:22metrics
- 15:07data
- 15:41FB Outrace
- 16:12Control Batch
- 16:49Application Latency
- 17:14Full Rollout
- 17:59Two Levels of Automation
- 19:22Raising Alarms
- 19:53Safety checks
- 20:43How did it roll out
- 21:22Fragmentation
- 22:48Future direction
- 23:43Any questions
related talks
52mHow ZNS SSD's Are Changing the Landscape
12mFusionRAID: Achieving Consistent Low Latency for Commodity SSD Arrays
45mBig Data and SSDs
53mOvercome Blast Radius Anxiety With Ultrafast Rebuilds and Performance for High-capacity SSDs
16mexF2FS: Transaction Support in Log-Structured Filesystem
42mCan File Systems Survive in Data-centric World?
47mEnabling Large Block Size devices in Linux
43mCeph and Flash