Profiling

This section covers CPU and memory profiling for df_engine.

CPU profiling (samply)

Requirements:

• samply

Installation:

cargo install --locked samply

Build:

cargo build --profile profiling --workspace

Run:

samply record ./target/profiling/df_engine --config ./configs/otap-noop.yaml

Result:

CPU profiling output will be automatically rendered on browser.

Memory profiling (dhat)

Build:

cargo build --profile profiling --no-default-features --features dhat-heap --workspace

Run:

./target/profiling/df_engine --config ./configs/otap-noop.yaml

Result:

On graceful shutdown of df_engine, it will generate dhat-heap.json file for memory profiling that needs to be rendered by uploading it to: https://nnethercote.github.io/dh_view/.

Live heap profiling (jemalloc pprof)

The admin server exposes a heap profiling endpoint at /api/v1/debug/pprof/heap that dumps unfreed heap allocations in pprof format. This requires jemalloc as the allocator (the default on Linux and macOS) with profiling enabled.

Run with profiling enabled:

_RJEM_MALLOC_CONF="prof:true,prof_active:true,lg_prof_sample:19" \
  cargo run -- --config ./configs/otap-noop.yaml

On Linux, use MALLOC_CONF instead of _RJEM_MALLOC_CONF.

Fetch a heap profile:

curl http://localhost:8080/api/v1/debug/pprof/heap -o heap.pprof

View with go tool pprof:

go tool pprof -http=:18080 ./target/debug/df_engine ./heap.pprof

Live CPU profiling (pprof)

The admin server exposes a CPU profiling endpoint at /api/v1/debug/pprof/profile that collects a CPU profile and returns it in pprof format. Not available on Windows.

Optional query parameters:

• seconds -- sampling duration in seconds (default 30).
• frequency -- sampling frequency in Hz (default 100).

Fetch a 10-second CPU profile:

curl "http://localhost:8080/api/v1/debug/pprof/profile?seconds=10" -o cpu.pprof

View with go tool pprof:

go tool pprof -http=:18080 ./target/debug/df_engine ./cpu.pprof