scope-profiler

This module provides a unified profiling system for Python applications, with optional integration of LIKWID markers using the pylikwid marker API for hardware performance counters.

It allows you to:

• Configure profiling globally via a singleton ProfilingConfig.
• Collect timing data via context-managed profiling regions.
• Use a clean decorator syntax to profile functions.
• Optionally record time traces in HDF5 files.
• Automatically initialize and close LIKWID markers only when needed, and store the resulting hardware counters and derived metrics in the same HDF5 file.
• Print aggregated summaries of all profiling regions.

Install

Install from PyPI:

pip install scope-profiler

Usage

To set up the configuration, create an instance of ProfilingConfig and add it to the ProfileManager, this should be done once at application startup and will persist until the program exits or is explicitly finalized (see below). Note that the config applies to any profiling contexts created (even in other files) after it has been initialized.

from scope_profiler import ProfileManager

# Setup global profiling configuration
ProfileManager.setup(
    use_likwid=False,
    recursive_profile=False,
)

# Profile the main() function with a decorator
@ProfileManager.profile("main")
def main():
    x = 0
    for i in range(10):
        # Profile each iteration with a context manager
        with ProfileManager.profile_region(region_name="iteration"):
            x += 1

# Call main
main()

# Finalize profiler
ProfileManager.finalize()

Execution:

❯ python test.py
profiling_data.h5  (1 rank(s))
  region     ranks  calls    total [s]      avg [s]
  --------------------------------------------------
  main           1      1   0.00150371   0.00150371
  iteration      1     10    3.832e-06    3.832e-07
  --------------------------------------------------
  TOTAL                11   0.00150754

finalize() prints the same table as scope-profiler inspect and ProfilingResults.print_summary(). Pass verbose=False to suppress it.

Inspecting a profiling file

scope-profiler inspect prints what is inside an HDF5 profiling file: the full run metadata (host, CPU, loaded modules, Slurm job, environment) and one statistics line per region, with no plotting dependencies needed.

scope-profiler inspect profiling_data.h5

==============================================================================
profiling_data.h5
2 rank(s), 4 region(s), 0.18 MiB, 0.0951538 s wall clock
==============================================================================

Metadata
  Run
    timestamp              : 2026-07-26T18:57:49
    user                   : mlindqvi
    hostname               : lrdn1234
  System
    chip_information       : AMD EPYC 9654 96-Core Processor
  Parallelism
    mpi_size               : 2
    omp_num_threads        : 8
    total_cores            : 16
  Slurm
    SLURM_JOB_ID           : 9988776
  Modules (4)
    profile/base
    gcc/12.3.0
    openmpi/4.1.6--gcc--12.3.0
    python/3.11.7

Regions (4)
  region    ranks  calls  total [s]    avg [s]
  ---------------------------------------------
  timestep      2      8   0.139235  0.0174044
  solve         2      8  0.0991292  0.0123911
  setup         2      2  0.0473326  0.0236663
  assemble      2      8  0.0399496  0.0049937
  ---------------------------------------------
  TOTAL               26   0.325647

Long values such as PATH are clipped unless --full is passed, regions can be filtered with --include/--exclude/--ranks, reordered with --sort, and either section shown alone with --metadata-only / --regions-only.

The metadata can also be exported to JSON, with one entry per inspected file and no clipping:

scope-profiler inspect profiling_data.h5 --export-metadata metadata.json --quiet

from scope_profiler.inspection import write_metadata_json

write_metadata_json("profiling_data.h5", "metadata.json")

Example plots

scope-profiler plot turns an HDF5 profiling file into Gantt, flame, duration, and speedup charts (see Flame graphs below for details). The plots here come from examples/generate_readme_figures.py, a small mock timestep loop with nested and self-recursive regions, and are saved to figures/:

python examples/generate_readme_figures.py

The flame graph for the same run is shown in Flame graphs below.

Overhead

The profiling overhead per call depends on the region type. The benchmark below (examples/benchmark_overhead.py) measures each mode against a bare function call:

The default TimeOnly mode — nanosecond timestamps for every call — adds roughly 0.33 µs per instrumented call.

Profiling can also be fully deactivated at setup time (deactivate_profiling=True) to reduce the overhead to ~0.1 µs — barely above a bare function call — making it safe to leave instrumentation in production code and toggle it on only when needed.

The LineProfiler mode is intentionally heavier (~50 µs/call) because line_profiler traces every source line. It is designed for targeted debugging of individual functions, not for always-on use in hot loops.

Profiling native code (C, C++, Fortran)

C and Fortran region APIs ship with the package, so native code — or the kernels under a Python driver — can be profiled into the same output. They share one trace format, so a program built from both lands in one profile.

#include "scope_profiler.h"

sp_init("profile", my_rank);
int solve = sp_region("solve");
sp_begin(solve);
solve_system();
sp_end(solve);
sp_finalize();

use scope_profiler
integer :: solve

call sp_init("profile", rank=my_rank)
solve = sp_region("solve")
call sp_begin(solve)
call solve_system()
call sp_end(solve)
call sp_finalize()

scope-profiler import-native . -o profiling_data.h5   # then plot/inspect as usual

Both are one self-contained file (Fortran 2008, or C99 with an extern "C" header for C++ callers): no HDF5, no MPI, nothing to link beyond libc.

Timestamps come from the same clock as Python's time.perf_counter_ns(), so a Python driver and the native kernels it calls land on a single timeline — ProfileManager.finalize(native_traces=".") folds them into one profile, with the native regions nested inside the Python ones that called them. See the Fortran and C guides.

Recursive profiling of nested calls

You can profile nested Python calls from one decorated entrypoint:

from scope_profiler import ProfileManager

ProfileManager.setup(recursive_profile=True)


def leaf(x):
    return x + 1


def inner(x):
    return leaf(x) * 2


@ProfileManager.profile("entry")
def entry():
    return sum(inner(i) for i in range(3))


entry()
ProfileManager.finalize()

When enabled, the profiler records regions for nested calls using fully qualified names (for example, my_module.inner), in addition to the main decorated region.

Zero-instrumentation CLI profiling

You can profile a whole script without touching its source, similar to python -m cProfile:

scope-profiler run my_script.py [script args...]
# equivalently: python -m scope_profiler run my_script.py [script args...]

Every Python function call the script makes is recorded as its own region under a name derived from its module and qualified name, using the same recursive tracer as recursive_profile=True above. By default only the script's own code is instrumented (the standard library and installed packages are skipped) to keep overhead low; pass --all to trace everything. Results are written to profiling_data.h5 by default (-o/--outfile to change it), and a per-region summary is printed unless -q/--quiet is given. Pass --line-profile to also persist line-by-line timings for the traced functions; this requires scope-profiler[line-profiler].

See examples/ex_cli_profiling.py for a script with no scope-profiler imports at all, run with:

scope-profiler run examples/ex_cli_profiling.py

Profiling self-recursive functions

A single region can also be safely re-entered by a recursive function - each call gets its own slot in the region's buffer, so nested calls don't overwrite each other's timing data. This works with both the decorator and context-manager forms:

from scope_profiler import ProfileManager

ProfileManager.setup()


@ProfileManager.profile("fibonacci")
def fibonacci(n):
    if n < 2:
        return n
    return fibonacci(n - 1) + fibonacci(n - 2)


def fibonacci_context_manager(n):
    with ProfileManager.profile_region("fibonacci_ctx"):
        if n < 2:
            return n
        return fibonacci_context_manager(n - 1) + fibonacci_context_manager(n - 2)


fibonacci(10)
fibonacci_context_manager(10)
ProfileManager.finalize()

Both fibonacci and fibonacci_ctx will report one call per recursive invocation, each with correct, non-overlapping timing data.

Analysing results in Python

read_h5() loads a merged profiling file into a ProfilingResults, which behaves like an ordered mapping of region name to region. Every duration and timestamp it reports is in seconds:

from scope_profiler import read_h5

results = read_h5("profiling_data.h5")
results.print_summary()

# region       calls     total [s]       avg [s]       min [s]       max [s]
# ---------------------------------------------------------------------------
# setup            1       0.02401       0.02401       0.02401       0.02401
# timestep         5      0.062835      0.012567     0.0087755     0.0187844

solve = results["solve"]          # an MPIRegion: the region across all ranks
solve.num_calls                   # summed over ranks
solve.total_duration              # seconds
solve.average_durations()         # {rank: seconds}, for load imbalance
solve[0].durations                # every call on rank 0, as a numpy array
solve.p50_duration                # median call duration, in seconds
solve.p95_duration                # 95th-percentile call duration
solve.rank_imbalance_pct          # slowest rank over mean, as a percentage

Regions can carry lightweight user-defined tags for downstream analysis:

with ProfileManager.profile_region("solve", tags=("compute", "hot")):
    solve()

results = ProfileManager.finalize(return_results=True)
results["solve"].tags  # ("compute", "hot")

summary() returns the same table as a list of dicts, and to_dataframe() returns it as a pandas DataFrame (one row per region, or per region and rank with per_rank=True):

frame = results.to_dataframe().sort_values("total_duration", ascending=False)
per_rank = results.to_dataframe(per_rank=True)

Summary rows and dataframes also include p50, p95, p99, and imbalance (the slowest rank's total time above the per-rank mean). These statistics are useful when averages hide tail latency or MPI load imbalance.

Safe profiling sessions

Use ProfileManager.session() when profiling should always be finalized, including when the profiled code raises:

with ProfileManager.session(file_path="run.h5", verbose=False,
                            return_results=True) as run:
    with ProfileManager.profile_region("solve"):
        solve()

results = run.results

include / exclude regexes select regions in get_regions(), summary(), to_dataframe() and every plot_* function.

Building your own plots

For custom analysis, work from the individual calls instead of the aggregates. events() returns one entry per recorded call, and to_events_dataframe() returns the same as a pandas DataFrame. Timestamps start at zero (the first region entry in the file), so they plot directly:

events = results.to_events_dataframe()
# columns: name, rank, call_index, start, end, duration   (seconds)

events.query("name == 'solve'")["duration"].hist(bins=50)
events.pivot_table(index="rank", columns="name", values="duration", aggfunc="sum")

Timestamps are measured from the start of the run, which setup() records. results.run_start_time is that instant, and results.startup_time the gap to the first profiled region — time the instrumentation never saw:

print(f"{results.startup_time:.3f} s before the first region was entered")

Files written without a start time (anything from before this existed) still read fine: run_start_time is then None, startup_time is 0.0, and the relative timeline falls back to the first region entry as before.

results.minimum_start_time, results.maximum_end_time and results.time_span bound the profiled window, and results.call_stack(rank=0) hands back the nesting the flame graph draws — one dict per call with depth and parent — so you can render your own nested view:

for call in results.call_stack(rank=0):
    print(f"{'  ' * call['depth']}{call['name']}: {call['duration']:.6f} s")

To post-process in the same script that recorded the data, use ProfileManager.read_results() after finalize() — it opens the file the current configuration wrote (on rank 0 under MPI).

The tutorial notebooks cover this in depth: getting started, post-processing, visualization, profiling modes, custom analysis and building your own plots.

Flame graphs

Because each call - including recursive re-entries of the same region - now has its own correctly nested (start, end) interval, the call stack can be reconstructed straight from the timing data and rendered as a flame graph, with recursion showing up as a narrowing tower of frames - as with refine_mesh below, from the same run shown in Example plots:

scope-profiler plot generates flame_plot.png alongside the Gantt chart for every run:

scope-profiler plot default profiling_data.h5 --show -o figures

Or programmatically:

from scope_profiler import read_h5, plot_flame

results = read_h5("profiling_data.h5")
plot_flame(results, filepath="flame_plot.png")

Gantt and flame charts (and plot_speedup) always color the same region the same way. Pass --cmap (or cmap= on the plot_* functions) to use a different matplotlib colormap than the default tab20:

scope-profiler plot default profiling_data.h5 --cmap viridis -o figures

By default the flame graph covers rank 0, since it represents a single execution's call stack; pass ranks=[...] to render one flame graph per requested rank.

Exporting plot data

Every plot_* function accepts a data_filepath argument that writes the exact data behind the chart to a file, so it can be re-parsed and re-plotted later without the original HDF5 file. data_format selects "csv" (default) or "json":

plot_gantt(results, filepath="gantt_plot.png", data_filepath="gantt_data.csv")
plot_gantt(
    results,
    filepath="gantt_plot.png",
    data_filepath="gantt_data.json",
    data_format="json",
)

The JSON payload additionally includes a colors map (region or file label to #rrggbb) matching the colors used in the matplotlib plot, so a JavaScript charting library like Plotly can reproduce the same look.

scope-profiler export plot-data does the same for selected plot kinds in one run, writing gantt_data, flame_data, durations_data, and (for multiple input files) speedup_data. Pass --format json to get .json files instead of the default .csv:

scope-profiler export plot-data profiling_data.h5 -o data
scope-profiler export plot-data profiling_data.h5 -o data --format json

Use --plots to restrict the exported data, useful when a website renders charts client-side (e.g. with Plotly) straight from the JSON:

scope-profiler export plot-data profiling_data.h5 -o data \
  --plots durations timeseries --format json

Viewing a run in snakeviz

scope-profiler export prof writes the profile in the .prof format of the standard library's cProfile, so a run can be explored with snakeviz or python -m pstats:

scope-profiler export prof profiling_data.h5 -o figures
snakeviz figures/profile_rank0.prof

Regions become "functions": cumtime is a region's total wall time and tottime is that minus the time spent in its nested regions. One file is written per exported rank, since .prof has no notion of ranks — see the CLI docs for the caveats of the reconstruction.

Viewing a run in speedscope

scope-profiler export speedscope writes the run as a speedscope JSON file. Unlike .prof, it keeps every individual call, so the timeline shows the run as it happened:

scope-profiler export speedscope profiling_data.h5 -o figures
npx speedscope figures/profile.speedscope.json  # or drop the file on speedscope.app

One file is written per input, holding one profile per exported rank, all sharing a time origin so ranks stay aligned. See the CLI docs for details.

MCP server for AI coding agents

pip install "scope-profiler[mcp]"
scope-profiler-mcp

scope-profiler-mcp exposes inspect_profile, compare_profiles, run_profile and plot_profile as MCP tools, so an agent such as Claude Code can inspect a run, benchmark a script, and check whether a code change made it faster or slower using structured data rather than parsed terminal output. It is a thin adapter over the same API described above -- see the MCP guide for installation, configuring Claude Code, and the full tool reference.