Skip to main content

Scalene: A high-resolution, low-overhead CPU and memory profiler for Python

Project description

scalene

scalene: a high-performance CPU and memory profiler for Python

by Emery Berger


About Scalene

Scalene is a high-performance CPU and memory profiler for Python that does a few things that other Python profilers do not and cannot do. It runs orders of magnitude faster than other profilers while delivering far more detailed information.

  1. 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).
  2. Scalene is precise. Unlike most other Python profilers, Scalene performs CPU profiling at the line level, pointing to the specific lines of code that are responsible for the execution time in your program. This level of detail can be much more useful information than the function-level profiles returned by most profilers.
  3. 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.

Installation

Scalene is distributed as a pip package. You can install it as follows:

  % pip install scalene

NOTE: Currently, installing Scalene in this way does not install its memory profiling library, so you will only be able to use it to perform CPU profiling. To take advantage of its memory profiling capability, you will need to download this repository.

Usage

The following command will run Scalene to only perform line-level CPU profiling on a provided example program.

  % python -m scalene test/testme.py

To perform both line-level CPU and memory profiling, you first need to build the specialized memory allocator by running make:

  % make

Profiling on a Mac OS X system:

  % DYLD_INSERT_LIBRARIES=$PWD/libscalene.dylib PYTHONMALLOC=malloc python -m scalene test/testme.py

Profiling on a Linux system:

  % LD_PRELOAD=$PWD/libscalene.so PYTHONMALLOC=malloc python -m scalene test/testme.py

Comparison to Other Profilers

Performance and Features

Below is a table comparing various profilers to scalene, running on an example Python program (benchmarks/julia1_nopil.py) from the book High Performance Python, by Gorelick and Ozsvald. All of these were run on a 2016 MacBook Pro.

Time (seconds) Slowdown Line-level? CPU? Memory? Unmodified code?
original program 7.76s 1.00x
cProfile 11.17s 1.44x function-level :heavy_check_mark: :heavy_check_mark:
Profile 278.19s 35.86x function-level :heavy_check_mark: :heavy_check_mark:
yappi 143.78s 18.53x function-level :heavy_check_mark: :heavy_check_mark:
line_profiler 93.27s 12.02x :heavy_check_mark: :heavy_check_mark: needs @profile decorators
memory_profiler aborted after 30 minutes >232x :heavy_check_mark: :heavy_check_mark: needs @profile decorators
scalene (CPU only) 8.31s 1.07x :heavy_check_mark: :heavy_check_mark: :heavy_check_mark:
scalene (CPU + memory) 9.11s 1.17x :heavy_check_mark: :heavy_check_mark: :heavy_check_mark: :heavy_check_mark:

Output

Scalene prints annotated source code for the program being profiled and any modules it uses in the same directory or subdirectories. Here is a snippet from pystone.py, just using CPU profiling:

benchmarks/pystone.py: % of CPU time =  88.34% out of   3.46s.
  Line	 | CPU %    |            benchmarks/pystone.py
  [... lines omitted ...]
   137	 |   0.65%  | 	def Proc1(PtrParIn):
   138	 |   1.96%  | 	    PtrParIn.PtrComp = NextRecord = PtrGlb.copy()
   139	 |   0.65%  | 	    PtrParIn.IntComp = 5
   140	 |   2.61%  | 	    NextRecord.IntComp = PtrParIn.IntComp
   141	 |   0.98%  | 	    NextRecord.PtrComp = PtrParIn.PtrComp
   142	 |   2.94%  | 	    NextRecord.PtrComp = Proc3(NextRecord.PtrComp)
   143	 |   0.65%  | 	    if NextRecord.Discr == Ident1:
   144	 |   0.33%  | 	        NextRecord.IntComp = 6
   145	 |   3.27%  | 	        NextRecord.EnumComp = Proc6(PtrParIn.EnumComp)
   146	 |   0.33%  | 	        NextRecord.PtrComp = PtrGlb.PtrComp
   147	 |   2.29%  | 	        NextRecord.IntComp = Proc7(NextRecord.IntComp, 10)
   148	 |          | 	    else:
   149	 |          | 	        PtrParIn = NextRecord.copy()
   150	 |   0.33%  | 	    NextRecord.PtrComp = None
   151	 |          | 	    return PtrParIn

And here is an example with memory profiling enabled, running the Julia benchmark.

benchmarks/julia1_nopil.py: % of CPU time =  96.23% out of   9.03s.
  Line	 | CPU %    | Memory (MB)| benchmarks/julia1_nopil.py
     1	 |          |         	 | 	# Pasted from Chapter 2, High Performance Python - O'Reilly Media;
     2	 |          |         	 | 	# minor modifications for Python 3 by Emery Berger
     3	 |          |         	 | 	
     4	 |          |         	 | 	"""Julia set generator without optional PIL-based image drawing"""
     5	 |          |         	 | 	import time
     6	 |          |         	 | 	# area of complex space to investigate
     7	 |          |         	 | 	x1, x2, y1, y2 = -1.8, 1.8, -1.8, 1.8
     8	 |          |         	 | 	c_real, c_imag = -0.62772, -.42193
     9	 |          |         	 | 	
    10	 |          |         	 | 	#@profile
    11	 |          |      0.12	 | 	def calculate_z_serial_purepython(maxiter, zs, cs):
    12	 |          |         	 | 	    """Calculate output list using Julia update rule"""
    13	 |          |      0.12	 | 	    output = [0] * len(zs)
    14	 |   0.35%  |     25.00	 | 	    for i in range(len(zs)):
    15	 |          |         	 | 	        n = 0
    16	 |   0.23%  |    -24.75	 | 	        z = zs[i]
    17	 |   0.46%  |    -27.75	 | 	        c = cs[i]
    18	 |  17.84%  |         	 | 	        while abs(z) < 2 and n < maxiter:
    19	 |  50.86%  |    998.12	 | 	            z = z * z + c
    20	 |  23.01%  |   -968.00	 | 	            n += 1
    21	 |   0.23%  |         	 | 	        output[i] = n
    22	 |          |         	 | 	    return output

Positive memory numbers indicate memory consumption in megabytes; negative memory numbers indicate memory reclamation. Note that because of the way Python's memory management works, frequent allocation and de-allocation (as in lines 19-20 above) show up as high positive memory on one line followed by an (approximately) corresponding negative memory on the following line(s).

Acknowledgements

Logo created by Sophia Berger.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for scalene, version 0.5.6
Filename, size File type Python version Upload date Hashes
Filename, size scalene-0.5.6-py3-none-any.whl (11.8 kB) File type Wheel Python version py3 Upload date Hashes View hashes
Filename, size scalene-0.5.6.tar.gz (7.9 kB) File type Source Python version None Upload date Hashes View hashes

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page