Using Pbench to debug Performance Problems
summary
Pbench packages an agent, server, dashboard to run benchmarks, start ~38 registered tools, capture six default configuration datasets, and archive runs with an md5 ID so results are reproducible; the team has stored 240,000+ tarballs (≈12.2 TB). Commands are register-toolset and user-benchmark which orchestrate SSH to many hosts, start/stop tools, postprocess perf/irq logs, build HTML graphs with jstart, and move results to a central server (no delete button yet).
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:03What problem does Pbench solve?
- 2:07History / Journey
- 5:32More Complex - uperf Example
- 6:51What is the Reference Point for Good?
- 10:26Solving the inconsistent and the incomplete
- 11:37Configuration Information
- 16:50pbench-register-tool-set
- 17:33pbench-user-benchmark - Details
- 20:52Other workloads
- 24:29Moving Data to the Pbench Server
- 27:36Pbench Dashboard - Controller
- 28:52Dashboard - Results Comparison
- 29:06Future Directions - Agent
- 30:45Future Directions - Dashboard
- 31:51Collected Data - Hierarchy (sample)
related talks
30mBuilding RailsPerf, a toolkit to detect performance regressions
21mPerformance Wins with BPF: Getting Started
17mStoring Continuous Benchmarking Data in Prometheus
35mPerformance Insights into eBPF Step by Step
08mFinding and Fixing Strange Builds
47mDive into BPF performance tools using python
37mPerformance Diagnostics with Open Distro for Elasticsearch
32mDog Food Can Be Delicious