Repository files navigation

byEmery Berger,Sam Stern, andJuan Altmayer Pizzorno.

Scalene community Slack

(tweet from Ian Ozsvald, author ofHigh Performance Python)

Scalene web-based user interface:http://plasma-umass.org/scalene-gui/

Scalene is a high-performance CPU, GPUand memory profiler forPython that does a number of things that other Python profilers do notand cannot do. It runs orders of magnitude faster than many otherprofilers while delivering far more detailed information. It is alsothe first profiler ever to incorporate AI-powered proposedoptimizations.

Note

To enable AI-powered optimization suggestions, you need to enter anOpenAI key in the box under "Advanced options".Your account will need to have a positive balance for this to work (check your balance athttps://platform.openai.com/account/usage).

Once you've entered your OpenAI key (see above), 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).

python3 -m pip install -U scalene

or

conda install -c conda-forge 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, installthe 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 theCommand 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.

scalene your_prog.py                             # full profile (outputs to web interface)python3 -m scalene your_prog.py                  # equivalent alternativescalene --cli your_prog.py                       # use the command-line only (no web interface)scalene --cpu your_prog.py                       # only profile CPUscalene --cpu --gpu your_prog.py                 # only profile CPU and GPUscalene --cpu --gpu --memory your_prog.py        # profile everything (same as no options)scalene --reduced-profile your_prog.py           # only profile lines with significant usagescalene --profile-interval 5.0 your_prog.py      # output a new profile every five secondsscalene (Scalene options) --- your_prog.py (...) # use --- to tell Scalene to ignore options after that pointscalene --help                                   # lists all options

fromscaleneimportscalene_profiler# Turn profiling onscalene_profiler.start()# your code# Turn profiling offscalene_profiler.stop()

fromscalene.scalene_profilerimportenable_profilingwithenable_profiling():# do something

# do not import profile!@profiledefslow_function():importtimetime.sleep(3)

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

By default, once Scalene has profiled your program, it will open atab in a web browser with an interactive user interface (all processing is donelocally). Hover over bars to see breakdowns of CPU and memoryconsumption, and click on underlined column headers to sort thecolumns. The generated fileprofile.html is self-contained and can be saved for later use.

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!

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

• Scalene performs profilingat the line levelandper function, pointing to the functions and the specific lines of code responsible for the execution time in your program.

• Scaleneseparates 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.
• Scalenehighlights hotspots (code accounting for significant percentages of CPU time or memory allocation) in red, making them even easier to spot.
• Scalene also separates outsystem time, making it easy to find I/O bottlenecks.

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

• Scaleneprofiles 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 ofmemory consumed by Python code vs. native code.
• Scalene producesper-line memory profiles.
• Scaleneidentifies lines with likely memory leaks.
• Scaleneprofilescopying volume, making it easy to spot inadvertent copying, especially due to crossing Python/library boundaries (e.g., accidentally convertingnumpy arrays into Python arrays, and vice versa).

• Scalene can producereduced 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 cansuspend and resume profiling.

Below is a table comparing theperformance 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 themultiprocessing 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

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 ituses in the same directory or subdirectories (you can optionally haveit--profile-all and only include files with at least a--cpu-percent-threshold of time). Here is a snippet frompystone.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").

The following command runs Scalene on a provided example program.

scalene test/testme.py

    % scalene --help     usage: scalene [-h] [--outfile OUTFILE] [--html] [--reduced-profile]                    [--profile-interval PROFILE_INTERVAL] [--cpu-only]                    [--profile-all] [--profile-only PROFILE_ONLY]                    [--use-virtual-time]                    [--cpu-percent-threshold CPU_PERCENT_THRESHOLD]                    [--cpu-sampling-rate CPU_SAMPLING_RATE]                    [--malloc-threshold MALLOC_THRESHOLD]     Scalene: a high-precision CPU and memory profiler.     https://github.com/plasma-umass/scalene     command-line:        % scalene [options] yourprogram.py     or        % python3 -m scalene [options] yourprogram.py     in Jupyter, line mode:        %scrun [options] statement     in Jupyter, cell mode:        %%scalene [options]        code...        code...     optional arguments:       -h, --help            show this help message and exit       --outfile OUTFILE     file to hold profiler output (default: stdout)       --html                output as HTML (default: text)       --reduced-profile     generate a reduced profile, with non-zero lines only (default: False)       --profile-interval PROFILE_INTERVAL                             output profiles every so many seconds (default: inf)       --cpu-only            only profile CPU time (default: profile CPU, memory, and copying)       --profile-all         profile all executed code, not just the target program (default: only the target program)       --profile-only PROFILE_ONLY                             profile only code in filenames that contain the given strings, separated by commas (default: no restrictions)       --use-virtual-time    measure only CPU time, not time spent in I/O or blocking (default: False)       --cpu-percent-threshold CPU_PERCENT_THRESHOLD                             only report profiles with at least this percent of CPU time (default: 1%)       --cpu-sampling-rate CPU_SAMPLING_RATE                             CPU sampling rate (default: every 0.01s)       --malloc-threshold MALLOC_THRESHOLD                             only report profiles with at least this many allocations (default: 100)     When running Scalene in the background, you can suspend/resume profiling     for the process ID that Scalene reports. For example:        % python3 -m scalene [options] yourprogram.py &      Scalene now profiling process 12345        to suspend profiling: python3 -m scalene.profile --off --pid 12345        to resume profiling:  python3 -m scalene.profile --on  --pid 12345

!pip install scalene%load_ext scalene

%scrun [options] statement

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

  % pip install -U scalene

  % python3 -m pip install -U scalene

  % sudo apt install git python3-all-dev

  % conda install -c conda-forge scalene

fromgeventimportmonkeymonkey.patch_all(thread=False)

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

@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}

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

Logo created bySophia Berger.

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