by Emery Berger
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.
- 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 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 than the function-level profiles returned by most profilers.
- Scalene separates out time spent running in Python from time spent 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 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 is distributed as a pip package and works on Linux and Mac OS X platforms. 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.
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
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? | Separates Python from native? | Memory? | Unmodified code? | |
|---|---|---|---|---|---|---|---|
| original program | 6.71s | 1.0x | |||||
cProfile |
11.04s | 1.65x | function-level | ✔️ | ✔️ | ||
Profile |
202.26s | 30.14x | function-level | ✔️ | ✔️ | ||
pyinstrument |
9.83s | 1.46x | function-level | ✔️ | ✔️ | ||
line_profiler |
78.0s | 11.62x | ✔️ | ✔️ | needs @profile decorators |
||
pprofile (deterministic) |
403.67s | 60.16x | ✔️ | ✔️ | ✔️ | ||
pprofile (statistical) |
7.47s | 1.11x | ✔️ | ✔️ | ✔️ | ||
yappi (CPU) |
127.53s | 19.01x | function-level | ✔️ | ✔️ | ||
yappi (wallclock) |
21.45s | 3.2x | function-level | ✔️ | ✔️ | ||
scalene (CPU only) |
6.98s | 1.04x | ✔️ | ✔️ | ✔️ | ✔️ | |
scalene (CPU + memory) |
7.68s | 1.14x | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
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 = 98.78% out of 3.47s.
| CPU % | CPU % |
Line | (Python) | (C) | [benchmarks/pystone.py]
--------------------------------------------------------------------------------
[... lines omitted ...]
137 | 0.87% | 0.13% | def Proc1(PtrParIn):
138 | 1.46% | 0.36% | PtrParIn.PtrComp = NextRecord = PtrGlb.copy()
139 | | | PtrParIn.IntComp = 5
140 | 0.87% | 0.04% | NextRecord.IntComp = PtrParIn.IntComp
141 | 1.46% | 0.30% | NextRecord.PtrComp = PtrParIn.PtrComp
142 | 2.33% | 0.26% | NextRecord.PtrComp = Proc3(NextRecord.PtrComp)
143 | 1.46% | -0.00% | if NextRecord.Discr == Ident1:
144 | 0.29% | 0.04% | NextRecord.IntComp = 6
145 | 1.75% | 0.40% | NextRecord.EnumComp = Proc6(PtrParIn.EnumComp)
146 | 1.75% | 0.29% | NextRecord.PtrComp = PtrGlb.PtrComp
147 | 0.58% | 0.12% | NextRecord.IntComp = Proc7(NextRecord.IntComp, 10)
148 | | | else:
149 | | | PtrParIn = NextRecord.copy()
150 | 0.87% | 0.15% | 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 = 99.22% out of 12.06s.
| CPU % | CPU % | Memory (MB) |
Line | (Python) | (C) | | [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 | | | | def calculate_z_serial_purepython(maxiter, zs, cs):
12 | | | | """Calculate output list using Julia update rule"""
13 | 0.08% | 0.02% | 0.06 | output = [0] * len(zs)
14 | 0.25% | 0.01% | 9.50 | for i in range(len(zs)):
15 | | | | n = 0
16 | 1.34% | 0.05% | -9.88 | z = zs[i]
17 | 0.50% | 0.01% | -8.44 | c = cs[i]
18 | 1.25% | 0.04% | | while abs(z) < 2 and n < maxiter:
19 | 68.67% | 2.27% | 42.50 | z = z * z + c
20 | 18.46% | 0.74% | -33.62 | n += 1
21 | | | | output[i] = n
22 | | | | return output
Positive memory numbers indicate total memory allocation 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).
Logo created by Sophia Berger.