mempulse

Tiny yet effective Python memory profiler/tracer. With minimized overhead with the C extension, mempulse gives you a holistic view of what's eating memory in your Python applications, on development and on production environments.

This package supports both Python 2.7 and Python >= 3.3 (tested with Python 3.13).

Install

Install from PyPI:

# Python-based `mempulse.MemoryUsageTracer` (depends on `psutil`, for Linux, macOS, Windows)
pip install mempulse[psutil]

# C-based `mempulse.cMemoryUsageTracer` (Linux only)
pip install mempulse

Install from source:

pip install setuptools
pip install .[psutil]

How to Use

Choose appropriate "tracing depth" (which slightly affects the overhead) then wrap the function you'd like to profile with a with statement:

import os
import sys
import mempulse

def application():
    callback = lambda r: sys.stderr.write(mempulse.format_trace_result(r))
    trace_depth = int(os.getenv('MEMORY_TRACE_DEPTH', '1')) # `0` disables tracing
    with mempulse.cMemoryUsageTracer(callback, trace_depth):
        the_workload()

Leavning the with scope will summarize line-by-line memory stats like belw:

Callsite         Method Name               USS   Swap       Peak RSS
--------------------------------------------------------------------
benchmark.py:53  run_mempulse_c     26,525,696      0     28,057,600
benchmark.py:29  workload           26,525,696      0     28,057,600
benchmark.py:30  workload           26,525,696      0     28,057,600
benchmark.py:31  workload           27,328,512      0     28,860,416
benchmark.py:32  workload          115,335,168      0    117,530,624
benchmark.py:33  workload          147,431,424      0    148,926,464
benchmark.py:34  workload          117,096,448      0    148,926,464
benchmark.py:35  workload           34,779,136      0    148,926,464
benchmark.py:36  workload           34,779,136      0    836,354,048
benchmark.py:37  workload           34,779,136      0    836,354,048
benchmark.py:38  workload           44,900,352      0    836,354,048
benchmark.py:39  workload           44,900,352      0    836,354,048
benchmark.py:52  run_mempulse_c     44,900,352      0    836,354,048

Benchmark

Running benchmark program examples/benchmark.py with Python 3.11 on macOS (with 8-core Intel Core i9 @ 3.6GHz) here showcases that mempulse outperforms similar tools such as memory_profiler, and tracemalloc, in terms of overhead:

Limitations

• Interoperability: mempulse sets trace function though sys.settrace() (or PyEval_SetStrace()), therefore it cannot be used with other tracers (such as Coverage.py) at the same time.
• Concurrency: mempulse traces current thread only. Functions run with other threads will not show in line-by-line trace records.

License

• This mempulse software is released under 3-Clause BSD License.
• The file mempulse/ext/uthash.h comes from uthash which is under BSD revised license.