PerfLens

The pipeline in one sentence: perf record → agent → TCP+zstd → server → parser → source mapper → SSE → browser. The agent runs perf record and perf stat in ~8-second rounds, flattens each round with perf script, compresses it 20–40× with zstd, and streams it over a 5-byte framed TCP protocol. The server parses samples, maps them back to source lines with addr2line, and pushes function tables, interactive SVG flame graphs, and heat-colored source to the browser over Server-Sent Events — no polling.

The constraints were the interesting part. Embedded targets don’t ship a modern Python, so the agent is written for Python 3.5 — no f-strings, no dataclasses — and there’s a single static C binary (~1.8 MB, vendored zstd) for targets with no Python at all. The server is pure stdlib, the UI is vanilla HTML/CSS/JS with hand-rolled flame graphs (no d3, no bundler). The agent probes perf_event_paranoid, enumerates which events the kernel actually supports, and tries call-graph modes (fp, dwarf, lbr) until one produces non-empty stacks — because on real devices, half of them don’t.

Cross-compilation is first-class: --toolchain-prefix derives the right addr2line and readelf, --sysroot resolves shared libraries and source trees, and the same agent code runs on aarch64, aarch64_be, armv7l, and x86_64. Sessions are saved as raw chunks and can be replayed lazily from the UI.