Accelerating the "slow" Python

/semut-io /Semutio /semut_io /semut.io

View Slide

Summary
Python is rapidly becoming the language of choice for many people, being
adopted in the Scientiﬁc Computing community.
The AI/Machine Learning community also loves the simplicity of the language,
but it all comes at a cost.
The question remains, whether or not we can bear the cost, or somehow
minimise it.
We will discuss 4-5 ways in which we can possibly minimise it.

View Slide

Outline
➔ Is Python really slow?
➔ “Why” is Python so slow?
➔ Why do I care about Python?
➔ What do I do when my code is slow?

View Slide

Who am I?
➔ SDE 1 @ Semut.io
➔ Electrical Engineer from IIT Mandi
➔ Managing Member @ PSF
➔ Lead Developer @ EinsteinPy
➔ Passionate about Black Holes and
Space (Scientiﬁc Computing)
➔ Love Mountains, and Himachal
Pradesh (India)

View Slide

Python is fast
Python vs. Java

View Slide

Python is slow
CPython has constant overhead per operation

View Slide

Python is slow

View Slide

Python is slow
Fortran is 100x faster for this simple task which computes ﬁbonacci
numbers.

View Slide

Can we get the best of both worlds?
➔ The simplicity and speed of writing Python Code
➔ Faster Execution

View Slide

Elephant in the room: Why is Python slow?
➔ Python is slower than C or Fortran for a variety of reasons.
➔ Some reasons are well known
➔ Some reasons are often ignored/ill-known
➔ Let’s dive right in...

View Slide

1. Python is dynamically typed
At the time of execution, the interpreter doesn’t know the type of variables
that are deﬁned.
Let’s just have a look at both Python and C codes for comparison, and
understand:

View Slide

C Addition
1. Assign 1 to a
○ 1a. Set a->PyObject_HEAD->typecode to
integer
○ 1b. Set a->val = 1
2. Assign 2 to b
○ 2a. Set b->PyObject_HEAD->typecode to
integer
○ 2b. Set b->val = 2
3. call binary_add(a, b)
○ 3a. ﬁnd typecode in a->PyObject_HEAD
○ 3b. a is an integer; value is a->val
○ 3c. ﬁnd typecode in b->PyObject_HEAD
○ 3d. b is an integer; value is b->val
○ 3e. call binary_add(a->val, b->val)
○ 3f. result of this is result, and is an
integer.
4. Create a Python object c
○ 4a. set c->PyObject_HEAD->typecode to
integer
○ 4b. set c->val to result
1. Assign 1 to a
2. Assign 2 to b
3. call binary_add(a, b)
4. Assign the result to c
Python Addition

View Slide

2. Python is interpreted, rather than compiled
➔ A smart compiler can look ahead and optimize for repeated and unneeded
operations.
➔ These can result in massive speed-ups.

View Slide

A lot of you may say,
“But Python is not as slow, I’ve used it.”

View Slide

3. Python’s object model can lead to inefficient
memory access
➔ While in C, you might use some buffer based array(pretty fast), Python’s
List is complicated, and clearly, slow.
➔ A NumPy array in its simplest form is a Python object build around a C
array. I.e. it has a pointer to a contiguous data buffer of values.
➔ A Python list, on the other hand, has a pointer to a contiguous buffer of
pointers, each of which points to a Python object which in turn has
references to its data (in this case, integers).
➔ It’s easy to see how numpy substantially boosts the performance.

View Slide

3. Python’s object model can lead to ineﬃcient
memory access

View Slide

Why am I even using Python?
➔ Dynamic typing makes Python easier to use than C
➔ It's extremely flexible and forgiving.
➔ This flexibility leads to efficient use of development time.
➔ Python offers easy hooks into compiled libraries, for when you need
that Fortranish performance. (See fortran-magic on PyPI)
➔ Python ends up being an extremely efficient language for the overall task
of doing science with code.
➔ Meta-Programming

View Slide

Example: K-Means Clustering
Algorithm:
1. Choose some Cluster Centers
2. Repeat:
a. Assign points to nearest center
b. Update center to mean of points
c. Check if Converged

View Slide

Pure Python Snippet
7.44 s ± 122 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) - 5000 points

View Slide

What can you do when Python is too slow?

View Slide

1. Line Proﬁling
“Premature optimization is the root of all evil”
~ Donald Knuth

View Slide

Repeated Operations on Arrays causing a lot of delay in computations.

View Slide

2. Numpy Vectorization (DSL)

View Slide

131 ms ± 3.68 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
Down from 7.44 seconds to 0.13 seconds (57X Faster)
Key: Repeated operations pushed into a compiled layer
Python overhead per array rather than per array element.

View Slide

Advantages:
➔ Python overhead per array rather than per array element
➔ Compact domain speciﬁc language for array operations
➔ NumPy is widely available
Disadvantages:
➔ Batch operations can lead to excessive memory usage
➔ Diﬀerent way of thinking about writing code
Recommendation: Use Numpy Everywhere

View Slide

3. Use specialised Data Structures
Some popular examples:
- Pandas
- Scipy
Same operations on Pandas Dataframe(groupby) and Scipy(KDTree):
102 ms ± 2.52 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
As compared to:
- 7.44 Seconds in Python
- 131 ms with NumPy

View Slide

Some interesting Data Structures
➔ scipy.spatial for spatial queries like distances, nearest neighbors, etc.
➔ pandas for SQL-like grouping & aggregation
➔ xarray for grouping across multiple dimensions
➔ scipy.sparse.csgraph for graph-like problems (e.g. ﬁnding shortest
paths)
Recommendation: Use whenever applicable and possible

View Slide

4. Cython
Few changes to the code, doesn’t quite look like Python. Steep Learning Curve.

View Slide

4. Cython
97.7 ms ± 12.2 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

View Slide

4. Cython
1.72 ms ± 27.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
56.8X Faster than Pure Python

View Slide

5. Numba JIT
➔ Numba is an open source, NumPy-aware optimizing compiler for Python
sponsored by Anaconda, Inc. It uses the LLVM compiler project to
generate machine code from Python syntax.
➔ GPU Support
➔ All you have to do is, add a @jit(nopython=True) decorator on function
you want to accelerate.
➔ Little bit buggy sometimes, but still works 99% of the time.

View Slide

5. Numba JIT

View Slide

5. Numba JIT
Pure Python: 97.7 ms ± 12.2 ms per loop (mean ± std. dev. of 7 runs, 10
loops each)
Numba JIT: 1.47 ms ± 14.2 µs per loop (mean ± std. dev. of 7 runs, 1000
loops each)
66X Performance Boost in Numba JIT
Again, this comes with a cost.

View Slide

Problems with using Numba JIT
➔ Two modes: “nopython” and “object”, only former is truly optimised.
➔ Functions JITted in nopython mode can only call functions JITted in
nopython mode.
➔ Avoid “object” mode. It is in process of being deprecated.
➔ Passing functions as arguments is even slower than JITing them, which is
a huge blocker.

View Slide

5. Numba JIT
High level API: Simple structure, but complex data structures. Can be
anything.
Dangerous Algorithms: Broken into small chunks which can easily be
accelerated using JIT. Avoid object mode.
Nice, High level API
Dangerous Algorithms

View Slide

5. Numba JIT
➔ If things stop working, export NUMBA_DISABLE_JIT=1
➔ You might have to rewrite some stuﬀ, makes the code a little less dynamic
➔ Numba evolves very fast, keep an eye on their webpage.

View Slide

Other ways of making Python faster
- CuPy
- Dask
There are no limits to what Python can do, thanks to the meta-programming
possible. Even you can write such a tool.
Recommendation: Use Numpy with Numba (Best for Scientiﬁc Computing)

View Slide

Resources
Pure Python Implementation:
https:/
/gist.github.com/shreyasbapat/bbc8f468dacbeb5b9baf7ead2c0d5077
Complete Notebook:
https:/
/gist.github.com/shreyasbapat/b7135d38273f72c6ce827256f4daedeb

View Slide

Thanks for joining!

View Slide

Accelerating the "slow" Python

Accelerating the "slow" Python

Shreyas Bapat

More Decks by Shreyas Bapat

Other Decks in Technology

Featured

Transcript