Investigate a change
Find which calls account for a cost change.
Collect a profile
From examples/bibliography/, with the controller prepared:
julia --startup-file=no --project=.controller/core run.jl profile_alloc Bibliography export_bibtex
julia --startup-file=no --project=.controller/core run.jl profile Bibliography export_bibtex
julia --startup-file=no --project=.controller/core run.jl wall_profile Bibliography export_bibtexKeep the printed report directory. Each command collects a separate profile of the same operation.
Read a flame graph
Width is the selected weight: CPU samples, task samples or allocation bytes.
Depth is nesting. Horizontal position is not a timeline.
A wide parent includes its descendants. Do not add parent and child percentages.
Follow a wide branch into the relevant function, then inspect its source line and full call path.
The same function can appear under different callers.
In the recorded CPU capture, export_bibtex accounts for about 98% of retained samples. Its _write_bibtex child is part of that share, not additional cost.
Check the effect of a change
Pick a call path that accounts for a large share.
Change the code (for export, avoid intermediate strings).
Rerun the original benchmark with the same input and verify the output.
Repeat when the difference is small relative to sample variation.
Profiling adds overhead, so confirm the change with BenchmarkTools or Chairmarks, not with the profile timings.
Other costs
Process memory — retained RAM.
Network measurement — traffic and packets.
Native profiling — Julia runtime and external libraries.
Recorded examples
Width is the share of retained CPU samples. Follow a wide export branch into its callees; the number of samples is neither an invocation count nor a duration in nanoseconds.
The full graph is fitted initially. Use + to enlarge its labels, then scroll inside the graph. Hover a rectangle, or use Tab to focus it, for its full stack and weight. Fit graph restores the overview.