DRBoost: Boosting Degraded Read Performance in MSR-Coded Storage Clusters
summary
DRBoost implements partial-chunk reconstruction, a reconstruction-friendly coding layout, and a fragmentation-free storage layout and integrates into Ceph. A testbed on Alibaba Cloud with 30 storage and 10 client nodes runs real traces from Alibaba, IBM, and Facebook. Results show degraded-read latency drops by 2.45–89.2× and IO amplification by 24.6–557×; per-technique gains are roughly 72.3×, 4.9×, and 1.2×, and normal reads see no performance loss.
AI-written summary · about
chapters (13)
- 0:00DRBoost and MSR-coded storage
- 1:22Erasure coding trade-offs
- 2:05MSR codes subpacketization
- 3:53Granularity mismatch in workloads
- 5:16Degraded read challenges
- 6:40DRBoost design decoupling
- 7:43Partial chunk reconstruction
- 9:30Reconstruction-friendly layouts
- 12:24Fragmentation-free layout
- 13:29Metadata and IO optimizations
- 14:13Evaluation setup and traces
- 15:16Performance improvements
- 16:21Conclusion and future
related talks
16mBoosting Full-Node Repair in Erasure-Coded Storage
17mParaRC: Embracing Sub-Packetization for Repair Parallelization in MSR-Coded Storage
35mMaking Ceph Fast in the Face of Failure
10mRevisiting Network Coding for Warm Blob Storage
12mDirect Access, High-Performance Memory Disaggregation with DirectCXL
38mEfficient Ceph Performance Troubleshooting in Production Using eBPF
34mModernizing Ceph Deployments at CERN: CephFS and Object Storage Across Data Centres
59mPractical Error Control Coding For Data Storage