Benchmarking Cloud Native Storage
summary
Benchmarks use sysbench, PGbench and CockroachDB workload tools on a 6-node 20-core, 120GB RAM, 2×200GB SSD cluster with a shared network. Tests compare bare‑metal filesystem, local volumes ~ bare metal (<1% overhead), and Rook/Ceph cloud‑native storage. Findings: Rook gives similar throughput but higher latency — roughly ~50% throughput on some DB workloads and ~2× latency for small random writes; sequential writes can be faster on Ceph due to parallelism. Recommendation: run multiple long runs and test on your hardware.
AI-written summary · about
chapters (8)
- 0:00Benchmarking Cloud Native Storage Josh Berkus, Red Hat
- 8:071. bare metal (no K8S) 2. node local storage (hostPath) 3. network storage (EBS)
- 18:036 blade cluster 20 cores ea. 128 GB RAM 2 SSDs w/ 200GB ea. shared network
- 19:59host filesystem • run tests using a host install, no Kubernetes gives reference numbers • using xfs & lvm
- 25:31network latency (1) used NodePort in order to run pgbench client on bare metal (2) extra network hops added command latency (3) pgbench sends a lot of short commands, with no batching
- 26:15rook storage • 5-node rook+ceph cluster • only 2 replicas (small cluster) • some default tweaks for performance • CockroachDB-on-Ceph, not Rook CDB
- 31:05Benchmark your own hardware with simple DB benchmarks to test your performance • Local Volume performance is equivalent to bare metal • Rook/Ceph has good throughput, but about double the latency for random writes
- 31:27Beware secondary issues that look like performance differences . On public cloud, cloud latency effects mask a lot of peformance differences
related talks
32mBenchmark Your Cloud Native Database
32mBenchmark Your Cloud Native Database
05mIgnite - How Fast Was My Kubernetes?
56mBenchmark Your Cloud-Native Database
31mRook: Intro and Deep Dive With Ceph
28mOpen Source Benchmarking for Ceph
47mThe Road Towards an Open Database Performance Ranking
31mFuture-proofing operational data at scale