.NET Performance Investigation
summary
Actual claim is measure not guess: use release builds, enable symbols and disable tiered JIT for small repros, prepare a narrow repro and collect thousands of samples (~4,000) before trusting results. Tools actually used: Visual Studio profiler, PerfView, dotnet-trace and BenchmarkDotNet plus speedcope and perf collect. Finds a managed vector math impl slower than native; a regression grows 250→306 ms; PerfView reports 3–5% overhead. Emphasizes tests, benchmarks and iterative profile->fix->verify.
AI-written summary · about
chapters (25)
- 0:00Intro
- 0:37No place for guessing
- 0:56Benchmark vs Profiler
- 1:16Recommended Settings
- 2:18Preparing the small repro case
- 3:34Personal recommendations
- 4:13Demo
- 9:11Paraview
- 10:13Paraview Tutorials
- 11:30Run Paraview
- 11:57Log
- 12:23CPU Stack
- 18:13Flame Graph
- 20:08Lining Methods
- 20:54Reading the Flame Graph
- 22:23Comparing apples to apples
- 24:46Regressions
- 27:15Dotnet Trace
- 34:38PerfCollect
- 37:25Architectural Changes
- 37:43Unit Tests
- 39:49Length Optimization
- 41:51Benchmarks
- 42:40summary
- 44:00Questions
related talks
01h 06mGet Better App Performance with this One Weird Trick
57mPragmatic Performance: When to care about perf, and what to do about it
01h 03mApplication Performance Optimisation in Practice
47mUnlocking Performance Improvements in .NET
46mHigh Performance .NET
37mRock Your Code Code and App Performance for Microsoft .Net
01h 00mRock Your Code: Code & App Performance for Microsoft .NET
01h 02mSqueezing the Hardware to Make Performance Juice