Understanding Ceph One Performance Counter at a Time
summary
Raw Ceph perf counters come straight from the daemon admin socket via perf dump and a supplied GitHub notebook with jq commands; an ad‑hoc web app extracts schemas. A single 4 MiB S3 PUT produces 18 ops into 3 messages over the messenger, about 4 KiB overhead, and replication shows one OSD receives the data then others get it via subops. The environment surfaces 8880 counters in ~41 groups; client op latency ~43 ms, OSD averages ~40 ms, while histograms reveal bimodal reads (sub‑ms and ~50 ms) and size×latency 2D buckets.
AI-written summary · about