scalene Documentation

File: README.md Copy Chapter

Scalene: a Python CPU+GPU+memory profiler with AI-powered optimization proposals

by Emery Berger, Sam Stern, and Juan Altmayer Pizzorno.

Scalene community Slack

(tweet from Ian Ozsvald, author of High Performance Python)

Python Profiler Links to AI to Improve Code Scalene identifies inefficiencies and asks GPT-4 for suggestions, IEEE Spectrum

Episode 172: Measuring Multiple Facets of Python Performance With Scalene, The Real Python podcast

Scalene web-based user interface: https://scalene-gui.github.io/scalene-gui/

About Scalene Copy Chapter

Scalene is a high-performance CPU, GPU and memory profiler for
Python that does a number of things that other Python profilers do not
and cannot do. It runs orders of magnitude faster than many other
profilers while delivering far more detailed information. It is also
the first profiler ever to incorporate AI-powered proposed
optimizations.

AI-powered optimization suggestions Copy Section

Note

For optimization suggestions, Scalene supports a variety of AI providers, including Amazon Bedrock, Microsoft Azure, OpenAI, and local models via Ollama. To enable AI-powered optimization suggestions from AI providers, you need to select a provider and, if needed, enter your credentials, in the box under "AI Optimization Options".

Once you've entered your key and any other needed data, click on the lightning bolt (⚡) beside any line or the explosion (💥) for an entire region of code to generate a proposed optimization. Click on a proposed optimization to copy it to the clipboard.

You can click as many times as you like on the lightning bolt or explosion, and it will generate different suggested optimizations. Your mileage may vary, but in some cases, the suggestions are quite impressive (e.g., order-of-magnitude improvements).

Quick Start Copy Section

Installing Scalene:

python3 -m pip install -U scalene

or

conda install -c conda-forge scalene

Using Scalene:

After installing Scalene, you can use Scalene at the command line, or as a Visual Studio Code extension.

Using the Scalene VS Code Extension:

First, install the Scalene extension from the VS Code Marketplace or by searching for it within VS Code by typing Command-Shift-X (Mac) or Ctrl-Shift-X (Windows). Once that's installed, click Command-Shift-P or Ctrl-Shift-P to open the Command Palette. Then select "Scalene: AI-powered profiling..." (you can start typing Scalene and it will pop up if it's installed). Run that and, assuming your code runs for at least a second, a Scalene profile will appear in a webview.

Commonly used command-line options:

Scalene uses a verb-based command structure with two main commands: run (to profile) and view (to display results).

# Profile a program (saves to scalene-profile.json)
scalene run your_prog.py
python3 -m scalene run your_prog.py              # equivalent alternative

# View a profile
scalene view                                     # open profile in browser
scalene view --cli                               # view in terminal
scalene view --html                              # save to scalene-profile.html
scalene view --standalone                        # save as self-contained HTML

# Common profiling options
scalene run --cpu-only your_prog.py              # only profile CPU (faster)
scalene run -o results.json your_prog.py         # custom output filename
scalene run -c config.yaml your_prog.py          # load options from config file

# Pass arguments to your program (use --- separator)
scalene run your_prog.py --- --arg1 --arg2

# Get help
scalene --help                                   # main help
scalene run --help                               # profiling options
scalene run --help-advanced                      # advanced profiling options
scalene view --help                              # viewing options

Using a YAML configuration file:

You can store Scalene options in a YAML configuration file and load them with -c or --config:

scalene run -c scalene.yaml your_prog.py

Example scalene.yaml:

# Output options
outfile: my-profile.json

# Profiling mode (use only one)
cpu-only: true              # CPU profiling only (faster)
# gpu: true                 # Include GPU profiling
# memory: true              # Include memory profiling

# Filter what gets profiled
profile-only: "mypackage,mymodule"    # Only profile these paths
profile-exclude: "tests,venv"          # Exclude these paths
profile-all: false                     # Profile all code, not just target

# Performance tuning
cpu-percent-threshold: 1     # Min CPU% to report (default: 1)
cpu-sampling-rate: 0.01      # Sampling interval in seconds
malloc-threshold: 100        # Min allocations to report

# Other options
use-virtual-time: false      # Measure CPU time only (not I/O)
stacks: true                 # Collect stack traces (default: true)
memory-leak-detector: true   # Detect likely memory leaks

Command-line arguments override config file settings.

Using Scalene programmatically in your code:

Invoke using scalene as above and then:

from scalene import scalene_profiler

# Turn profiling on
scalene_profiler.start()

# your code

# Turn profiling off
scalene_profiler.stop()

from scalene.scalene_profiler import enable_profiling

with enable_profiling():
    # do something

Using Scalene to profile only specific functions via @profile:

Just preface any functions you want to profile with the @profile decorator and run it with Scalene:

# do not import profile!

@profile
def slow_function():
    import time
    time.sleep(3)

Web-based GUI

Scalene has both a CLI and a web-based GUI (demo here).

By default, once Scalene has profiled your program, it will open a
tab in a web browser with an interactive user interface (all processing is done
locally). Hover over bars to see breakdowns of CPU and memory
consumption, and click on underlined column headers to sort the
columns. The GUI works fully offline with no internet connection required.

Use scalene view --standalone to generate a completely self-contained HTML file with all assets embedded, perfect for sharing or archiving.

Scalene Overview Copy Chapter

Scalene talk (PyCon US 2021) Copy Section

This talk presented at PyCon 2021 walks through Scalene's advantages and how to use it to debug the performance of an application (and provides some technical details on its internals). We highly recommend watching this video!

Fast and Accurate Copy Section

• Scalene is fast. It uses sampling instead of instrumentation or relying on Python's tracing facilities. Its overhead is typically no more than 10-20% (and often less).

• Scalene is accurate. We tested CPU profiler accuracy and found that Scalene is among the most accurate profilers, correctly measuring time taken.

• Scalene performs profiling at the line level and per function, pointing to the functions and the specific lines of code responsible for the execution time in your program.

CPU profiling Copy Section

• Scalene separates out time spent in Python from time in native code (including libraries). Most Python programmers aren't going to optimize the performance of native code (which is usually either in the Python implementation or external libraries), so this helps developers focus their optimization efforts on the code they can actually improve.
• Scalene highlights hotspots (code accounting for significant percentages of CPU time or memory allocation) in red, making them even easier to spot.
• Scalene also separates out system time, making it easy to find I/O bottlenecks.

GPU profiling Copy Section

• Scalene reports GPU time (currently limited to NVIDIA-based systems).

Memory profiling Copy Section

• Scalene profiles memory usage. In addition to tracking CPU usage, Scalene also points to the specific lines of code responsible for memory growth. It accomplishes this via an included specialized memory allocator.
• Scalene separates out the percentage of memory consumed by Python code vs. native code.
• Scalene produces per-line memory profiles.
• Scalene identifies lines with likely memory leaks.
• Scalene profiles copying volume, making it easy to spot inadvertent copying, especially due to crossing Python/library boundaries (e.g., accidentally converting numpy arrays into Python arrays, and vice versa).

Async profiling Copy Section

Scalene attributes wall-clock await time to the exact line where the
coroutine is suspended, alongside the mean and peak number of
coroutines waiting concurrently on that line. This makes it easy to spot
serialized I/O (long await, low concurrency) versus genuinely-parallel
fan-out (long await, high concurrency). The await column in the GUI
draws a clockwise pie wedge per line; bigger wedges mean more time spent
suspended on that await.

Async profiling is on by default; pass --no-async to disable it.
Overhead on non-async code is negligible (one frozenset lookup per
signal, and sys.monitoring callbacks never fire when no coroutines
are running).

Stack views (call stacks, memory flame chart, timeline) Copy Section

When Scalene is run with --stacks (the default — pass --no-stacks
to disable), it captures stitched Python + native (C/C++) stacks at
every CPU and allocation sample, and presents them three ways in the
GUI:

• Call stacks — a flame chart of the top stitched stacks across the
  whole run. Python frames are clickable to jump to source. Native
  frames are demangled and shown alongside Python frames in the same
  picture, so you can see exactly which library call is on the hot path.

• Memory stacks — a memory-weighted flame chart where frame width is
  proportional to MB allocated along that call path (not sample
  count). Useful for finding which call paths are responsible for a
  program's memory footprint, especially when allocations come from deep
  inside a library.

• Timeline — an icicle/time-series view of the same stitched stacks
  laid out left-to-right by wall-clock time, with tracks at the top
  marking GC, I/O, and GIL activity. Coroutines appear as [await] task_name frames so you can see when each task ran versus waited.

All three views are collapsible and resizable; expand them via the
disclosure triangles at the bottom of the GUI.

Other features Copy Section

• Scalene can produce reduced profiles (via --reduced-profile) that only report lines that consume more than 1% of CPU or perform at least 100 allocations.
• Scalene supports @profile decorators to profile only specific functions.
• When Scalene is profiling a program launched in the background (via &), you can suspend and resume profiling.
• Scalene supports free-threaded Python (3.13t / 3.14t) with full
  CPU + memory profiling.
• Scalene's GUI is fully offline / air-gapped — assets are vendored
  and scalene view --standalone produces a single self-contained HTML
  file you can email or archive.

Comparison to Other Profilers

Performance and Features Copy Chapter

Below is a table comparing the performance and features of various profilers to Scalene.

• Slowdown: the slowdown when running a benchmark from the Pyperformance suite. Green means less than 2x overhead. Scalene's overhead is just a 35% slowdown.

Scalene has all of the following features, many of which only Scalene supports:

• Lines or functions: does the profiler report information only for entire functions, or for every line -- Scalene does both.
• Unmodified Code: works on unmodified code.
• Threads: supports Python threads.
• Multiprocessing: supports use of the multiprocessing library -- Scalene only
• Python vs. C time: breaks out time spent in Python vs. native code (e.g., libraries) -- Scalene only
• System time: breaks out system time (e.g., sleeping or performing I/O) -- Scalene only
• Profiles memory: reports memory consumption per line / function
• GPU: reports time spent on an NVIDIA GPU (if present) -- Scalene only
• Memory trends: reports memory use over time per line / function -- Scalene only
• Copy volume: reports megabytes being copied per second -- Scalene only
• Detects leaks: automatically pinpoints lines responsible for likely memory leaks -- Scalene only
• Async / await profiling: per-line wall-clock attribution of suspended coroutine time, with mean and peak concurrency -- Scalene only
• Stitched Python + native stacks: flame charts and a wall-clock timeline that combine Python and C/C++ frames in one picture -- Scalene only

Output Copy Chapter

If you include the --cli option, Scalene prints annotated source code for the program being profiled
(as text, JSON (--json), or HTML (--html)) and any modules it
uses in the same directory or subdirectories (you can optionally have
it --profile-all and only include files with at least a
--cpu-percent-threshold of time). Here is a snippet from
pystone.py.

• Memory usage at the top: Visualized by "sparklines", memory consumption over the runtime of the profiled code.
• "Time Python": How much time was spent in Python code.
• "native": How much time was spent in non-Python code (e.g., libraries written in C/C++).
• "system": How much time was spent in the system (e.g., I/O).
• "GPU": (not shown here) How much time spent on the GPU, if your system has an NVIDIA GPU installed.
• "Memory Python": How much of the memory allocation happened on the Python side of the code, as opposed to in non-Python code (e.g., libraries written in C/C++).
• "net": Positive net memory numbers indicate total memory allocation in megabytes; negative net memory numbers indicate memory reclamation.
• "timeline / %": Visualized by "sparklines", memory consumption generated by this line over the program runtime, and the percentages of total memory activity this line represents.
• "Copy (MB/s)": The amount of megabytes being copied per second (see "About Scalene").

Scalene Copy Chapter

The following command runs Scalene on a provided example program.

scalene test/testme.py

Click to see all Scalene's options (available by running with --help)

/* Detailed source-code truncated for AI context efficiency. */

Scalene with Jupyter Copy Section

Instructions for installing and using Scalene with Jupyter notebooks

This notebook illustrates the use of Scalene in Jupyter.

Installation:

!pip install scalene
%load_ext scalene

Line mode:

%scrun [options] statement

Cell mode:

%%scalene [options]
code...
code...

Installation Copy Chapter

Using pip (Mac OS X, Linux, Windows, and WSL2)

Scalene is distributed as a pip package and works on Mac OS X, Linux (including Ubuntu in Windows WSL2) and Windows platforms.

Note for Windows users

Starting with Scalene 2.0, Windows supports full memory profiling. If you
encounter issues, ensure you have the Visual C++ Redistributable
installed. If building from source, you will need Visual C++ Build Tools and CMake.

You can install it as follows:

% pip install -U scalene

or

% python3 -m pip install -U scalene

You may need to install some packages first.

See https://stackoverflow.com/a/19344978/4954434 for full instructions for all Linux flavors.

For Ubuntu/Debian:

% sudo apt install git python3-all-dev

Using conda (Mac OS X, Linux, Windows, and WSL2)

% conda install -c conda-forge scalene

Scalene is distributed as a conda package and works on Mac OS X, Linux (including Ubuntu in Windows WSL2) and Windows platforms.

Note for Windows users

Starting with Scalene 2.0, Windows supports full memory profiling. If you
encounter issues, ensure you have the Visual C++ Redistributable
installed.

On ArchLinux

You can install Scalene on Arch Linux via the AUR
package. Use your favorite AUR helper, or
manually download the PKGBUILD and run makepkg -cirs to build. Note that this will place
libscalene.so in /usr/lib; modify the below usage instructions accordingly.

Frequently Asked Questions

Can I use Scalene with PyTest?

A: Yes! You can run it as follows (for example):

scalene run -m pytest your_test.py

or

python3 -m scalene run -m pytest your_test.py

Is there any way to get shorter profiles or do more targeted profiling?

A: Yes! There are several options:

1. Use --reduced-profile to include only lines and files with memory/CPU/GPU activity.
2. Use --profile-only to include only filenames containing specific strings (as in, --profile-only foo,bar,baz).
3. Decorate functions of interest with @profile to have Scalene report only those functions.
4. Turn profiling on and off programmatically by importing Scalene profiler (from scalene import scalene_profiler) and then turning profiling on and off via scalene_profiler.start() and scalene_profiler.stop(). By default, Scalene runs with profiling on, so to delay profiling until desired, use the --off command-line option (scalene run --off yourprogram.py).

How do I run Scalene in PyCharm?

A: In PyCharm, you can run Scalene at the command line by opening the terminal at the bottom of the IDE and running a Scalene command (e.g., scalene run <your program>). Then use scalene view --html to generate an HTML file (scalene-profile.html) that you can view in the IDE.

How do I use Scalene with Django?

A: Pass in the --noreload option (see https://github.com/plasma-umass/scalene/issues/178).

Does Scalene work with gevent/Greenlets?

A: Yes! Put the following code in the beginning of your program, or modify the call to monkey.patch_all as below:

from gevent import monkey
monkey.patch_all(thread=False)

How do I use Scalene with PyTorch on the Mac?

A: Scalene works with PyTorch version 1.5.1 on Mac OS X. There's a bug in newer versions of PyTorch (https://github.com/pytorch/pytorch/issues/57185) that interferes with Scalene (discussion here: https://github.com/plasma-umass/scalene/issues/110), but only on Macs.

Technical Information

For details about how Scalene works, please see the following paper, which won the Jay Lepreau Best Paper Award at OSDI 2023: Triangulating Python Performance Issues with Scalene. (Note that this paper does not include information about the AI-driven proposed optimizations.)

To cite Scalene in an academic paper, please use the following:

@inproceedings{288540,
author = {Emery D. Berger and Sam Stern and Juan Altmayer Pizzorno},
title = {Triangulating Python Performance Issues with {S}calene},
booktitle = {{17th USENIX Symposium on Operating Systems Design and Implementation (OSDI 23)}},
year = {2023},
isbn = {978-1-939133-34-2},
address = {Boston, MA},
pages = {51--64},
url = {https://www.usenix.org/conference/osdi23/presentation/berger},
publisher = {USENIX Association},
month = jul
}

Success Stories

If you use Scalene to successfully debug a performance problem, please add a comment to this issue!

Acknowledgements

Logo created by Sophia Berger.

This material is based upon work supported by the National Science
Foundation under Grant No. 1955610. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of
the author(s) and do not necessarily reflect the views of the National
Science Foundation.

File: scalene/scalene-gui/README.md Copy Chapter

This repository holds the Scalene GUI.

To use the interactive demo, go to the Scalene GUI site.

Acknowledgements

Scalene's web GUI makes use of the following wonderful systems. Many thanks to their authors.

• Vega-Lite for rendering graphs
• Prism for code highlighting
• tablesort for sorting tables by columns

This material is based upon work supported by the National Science
Foundation under Grant No. 1955610. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of
the author(s) and do not necessarily reflect the views of the National
Science Foundation.

File: scalene/README.md Copy Chapter

Helper functions: Copy Section

• adaptive.py:
  Adaptive maintains samples used for memory footprint sparklines.

• profile.py:
  used to suspend/resume profiling of Scalene when run in the background (with &).

• replacement_*.py:
  Scalene needs to interpose on a variety of functions; this functionality is in these files.

• runningstats.py:
  RunningStats takes input samples and incrementally computes average and other statistics in a fixed amount of space.

• sparkline.py:
  Functions to generate sparklines, used to display memory consumption over time.

• syntaxline.py:
  A helper function for pretty-printing with the Rich library.

Core Scalene functions: Copy Section

• scalene_arguments.py:
  ScaleneArguments holds command-line arguments and their default values.

• scalene_gpu.py:
  ScaleneGPU wraps the NVIDIA library to conveniently provide access to GPU statistics required by Scalene.

• scalene_magics.py:
  Sets up the "magics" for using Scalene within Jupyter notebooks (%scrun and %%scalene).

• scalene_output.py:
  ScaleneOutput encapsulates functions used for generating Scalene's profiles either as text or HTML.

• scalene_profiler.py:
  The core of the Scalene profiler.

• scalene_signals.py:
  Defines the Unix signals that Scalene uses (some of which must be kept in sync with include/sampleheap.hpp).

• scalene_statistics.py:
  Operations for managing the statistics generated by Scalene.

• scalene_version.py:
  The version number of Scalene which ultimately is reflected on pypi (for pip installs, used by setup.py in the top level directory).