P PmUserZFS Well-Known Member Feb 2, 2018 141 7 58 Today at 14:30 #1 I ask copilot for some performance numbers I got this as ansser Pool TypeWrite LatencyRead LatencyReplicated~1–2 ms (SSD + network RTT)~1 msEC 4+2~3–5 ms (encode + 6 OSD commit)~2–4 ms (decode) Is EC that bad?
I ask copilot for some performance numbers I got this as ansser Pool TypeWrite LatencyRead LatencyReplicated~1–2 ms (SSD + network RTT)~1 msEC 4+2~3–5 ms (encode + 6 OSD commit)~2–4 ms (decode) Is EC that bad?
P PmUserZFS Well-Known Member Feb 2, 2018 141 7 58 Today at 14:31 #2 Replicated size=3 NVMe pool: Writes: ~0.5–1 ms (ACK after 2 copies). Reads: ~0.2–0.5 ms. EC NVMe pool: Writes: ~2–4 ms. Reads: ~1–2 ms. So EC on NVMe is still fast, but replication will always beat EC for latency-sensitive workloads because: EC requires full stripe commit. Extra CPU and network steps.
Replicated size=3 NVMe pool: Writes: ~0.5–1 ms (ACK after 2 copies). Reads: ~0.2–0.5 ms. EC NVMe pool: Writes: ~2–4 ms. Reads: ~1–2 ms. So EC on NVMe is still fast, but replication will always beat EC for latency-sensitive workloads because: EC requires full stripe commit. Extra CPU and network steps.
C cfgmgr Member Jul 25, 2023 64 13 13 Today at 18:27 #3 Theres a bunch of EC changes coming up in Ceph 20. Worth a read. Of course always dependent on your workload. https://ceph.io/en/news/blog/2025/tentacle-fastec-performance-updates/
Theres a bunch of EC changes coming up in Ceph 20. Worth a read. Of course always dependent on your workload. https://ceph.io/en/news/blog/2025/tentacle-fastec-performance-updates/
