Evaluating the Performance and Power Efficiency of Ceph Storage Configurations
summary
Benchmarks compare three 24‑drive node types (8 TB SATA, 22 TB dense HDD, 15 TB TLC NVMe) using RADOS benchmarking across EC layouts (notably 4+2 EC and 8+3 EC). Tests use ~2 PB per configuration and show streaming GB/s per usable PB is similar across HDD and NVMe, but NVMe delivers that throughput with lower watts per GB/s and vastly higher small‑IO IOPS. Denser 22 TB drives cut random read IOPS per TB; idle power dominates (150 PB cluster ≈100 kW idle, +~20 kW at full read). Guidance: use the densest HDDs that meet capacity+performance, otherwise choose dense NVMe. RADOS benchmarking 15 TB TLC NVMe
AI-written summary · about
chapters (10)
- 0:00Introduction and context
- 3:31Storage needs for scientific computing
- 6:03Test configurations and hardware
- 8:08Streaming IO performance analysis
- 13:01Small IO performance comparison
- 16:57Power usage and efficiency
- 23:42Power efficiency summary
- 26:30Idle power dominance and tape archives
- 27:35Recommendations for storage procurement
- 30:04Q&A session
related talks
40mBenchmarking: Repeatable & Comparable
31mAI, ML, and the Ceph Advantage: Scalable Storage for Smarter Workflows
36mStorage Wars
29mLarge-Scale Benchmark Results of NVMe Over Fabric Technology
43mComparing Performance of NVMe Hard Drives in KVM, Baremetal, and Docker
25mPerformance Analysis of NVMe SSDs and their Implication on Real World Databases
43mCeph and Flash
39mBenchmarking Cloud Native Storage