Made some progress with my low-latency C++ setup. Now has 2 environments:
- a mobile “dev/test”: Ubuntu VM running on a MacBook
- a fixed “production”: bare metal Linux on a 2013 MacBbook with Intel Core i7-3635QM (quad-core Ivy Bridge chip from 2012)
Claude Code is running on both.
👉the “dev” machine is where coding and initial testing happens
- code, compile and build in vscode
- test algorithms for performance and correctness with a synthetic traffic generator.
- measure with Google Benchmark
👉the “prod” machine for more “realistic” tests:
- for instance: feed binaries with actual market data traffic (from NASDAQ historical market data sample file)
- Because it’s an actual Intel machine and not a VM, a number of things are possible:
- inspect hardware counters, perform TMA investigation (see screenshot).
- perf stat can display TopdownL1, TopdownL2, TopdownL3 and all tma_xxx counters.
- isolate cores
- pin threads to cores
- enable/disable huge pages
- disable Turboboost, disable hyper-threading,
- build flame graphs
“prod” stays at home because it’s an old Macbook from 2013 whose swelling battery was surgically removed.
“prod” can be remote-controlled from “dev” through the Claude session (or from SSH / remote desktop potentially)
