Is writing performant code too expensive?
summary
Argues that targeted optimization can beat big-cloud defaults with real numbers and experiments: uses JMH microbenchmarks and a NUMA ping‑pong test to show latency from ~200ns to ~650ns depending on placement, measures ~18 GB/s per core streaming memory bandwidth dropping to ~400 MB/s or lower for random access, shows a simple branchy integer scan runs 43ms or 500ms depending on branch prediction while a branchless/vectorized version reaches ~11ms, and cites the COST metric showing single‑thread C often outperforms large clusters.
AI-written summary · about
chapters (15)
- 0:00Introduction and welcome
- 1:33Performance obsession in coding
- 8:36Startup greenhouse failure story
- 11:08Validating ideas quickly
- 14:02Cost of engineering vs servers
- 17:42Hardware progress and inefficiency
- 19:10Zero point one developer mindset
- 20:21Research into data processing
- 23:17Benchmarking old database systems
- 25:28Streaming systems scalability myth
- 29:06NUMA architecture basics
- 32:23Hardware optimization in Java
- 43:46Branch prediction and performance
- 48:29Balancing optimization approaches
- 49:11Q&A and closing remarks