Linux Systems Performance
summary
An application shows a claimed 3x speedup; root cause analysis proves it: IPC 0.56 vs 1.5 (container host runs three times more instructions per cycle), LLC hit ratio 30%→90%, and context switches 2,000,000 vs 1,000. Analysis uses vmstat/mpstat/iostat, perf and hardware PMCs (pmcarch), BPF tools (bcc, bpftrace), cpudist for CPU slice times, and flame graphs/flame scope for visualization. Custom tools and methods identify tenant contention, cache warming time (µs vs ms), CPU type differences, and the speedup is explained and reproducible.
AI-written summary · about
chapters (34)
- 0:00Introduction
- 0:53NBStat
- 2:35PMC Arch
- 6:28Curve
- 7:12CP dist
- 8:28Systems Performance
- 13:06Load Averages
- 16:48Top
- 17:07Htop
- 18:27VMStat
- 20:47Free
- 21:32Perf
- 23:32TCP Dump
- 24:40Netstat
- 25:20SS Slabtop
- 25:51Containers
- 26:04Show Boost
- 26:57Static Performance Tuning
- 27:25Methodology
- 27:45Linux Performance Analysis
- 32:38Profiling
- 33:32Flame graphs
- 34:13BPF
- 34:31Flamescope
- 34:55Perfect Profile
- 35:23Tracing
- 35:41Tracing Stack
- 36:00Trace
- 36:48HD for slower
- 37:23File System
- 38:13CPU Analysis
- 38:30Netflix Tuning
- 39:07Queue Discs
- 39:39Summary
related talks
51mLinux Systems Performance - Architecture
48mFantastic Performance and Where to Find It
30m"Logs Told Us It Was Kernel – It Wasn't"
35mOpening the Box: Diagnosing Operating-System Task-Scheduler...
45mPerformance lab. The power of the perf tool!
26mLatency Distributions and Micro-Benchmarking to Identify and Characterize Kernel Hotspots
27mDebugging Linux Issues with eBPF
46mJava vs CPU