Speeding things up
Contents
Speeding things up¶
Slow code is nothing to be ashamed of. If the program runs and produces correct results, the important work is done. But when the time does come to speed things up, it can be difficult to know where to start. Don’t panic and write a blog post about how slow Python/R/Julia is, instead try to utilize some of these tools and strategies.
Measuring performance¶
To speed up slow code, the first step is to determine what is slowing things down. For simple programs it may be easy to intuitively identify bottlenecks. Trusting your gut ceases to be an appropriate strategy as codebases grow in size and complexity.
Measurement tools such as profilers and timers are invaluable allies when optimizing programs. Explicitly measuring the speed of code execution can shine light on subtle bottlenecks and inefficiencies which may not have been discovered otherwise.
Timers¶
Timers measure the time it takes to execute an expression. Using a timer helps to isolate and test the speed of a small snippet or function. The code block below demonstrates using the IPython %timeit magic command to compare two different ways to build a list of squares in Python:
In \[1\]: def comprehension(n):
...: return \[i\*\*2 for i in range(n)\]
...:
In \[2\]: def loop(n):
...: x = \[\]
...: for i in range(n):
...: x.append(i\*\*2)
...: return x
...:
In \[3\]: %timeit comprehension(1000)
190 µs ± 1.35 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
In \[4\]: %timeit loop(1000)
226 µs ± 2.44 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Profilers¶
Profilers measure how often different parts of a program are called and how long they take to execute. For example, consider the following inefficient Python program:
\# fib_buzz.py
def fib(n):
if n < 2:
return n
else:
return fib(n - 1) + fib(n - 2)
def fib_buzz(n): for i in range(n): if fib(i) % 3 == 0 and fib(i) % 5 == 0: print(“fibbuzz”)
elif fib(i) % 3 == 0:
print("fib")
elif fib(i) % 5 == 0:
print("buzz")
else:
print(i)
Profiling fib_buzz with the IPython %prun magic command produces the following table:
In \[1\] %load fib_buzz.py
In \[2\] %prun -s calls -l 10 fib_buzz(30)
13048500 function calls (722 primitive calls) in 2.959 seconds
Ordered by: call count
List reduced from 15 to 10 due to restriction <10>
ncalls tottime percall cumtime percall filename:lineno(function)
13047866/88 2.948 0.000 2.948 0.034 <ipython-input-3-a0fe65949268>:2(fib)
120 0.000 0.000 0.000 0.000 {method 'finditer' of 're.Pattern' objects}
60 0.000 0.000 0.010 0.000 ansitowin32.py:40(write)
60 0.000 0.000 0.010 0.000 ansitowin32.py:160(write)
60 0.000 0.000 0.010 0.000 ansitowin32.py:177(write_and_convert)
60 0.000 0.000 0.009 0.000 ansitowin32.py:193(write_plain_text)
60 0.000 0.000 0.000 0.000 ansitowin32.py:245(convert_osc)
60 0.004 0.000 0.004 0.000 {method 'write' of '_io.TextIOWrapper' objects}
60 0.006 0.000 0.006 0.000 {method 'flush' of '_io.TextIOWrapper' objects}
60 0.000 0.000 0.000 0.000 {built-in method builtins.len}
This output shows that fib function calls make up the majority of the program runtime. Optimization efforts should be focused on the fib function, either by reducing the number of times it is called or by speeding up execution time per call.
Tips for writing faster code¶
Writing fast code to solve problems in scientific computing is a lot like writing fast code in any other domain. However, there are some specific performance considerations to keep in mind, especially regarding n-dimensional arrays. Below are a few tips for writing faster scientific code:
1. Use vectorized operations¶
When working with data stored in arrays, vectorized functions are much faster than using a loop to apply an operation element by element. Vectorized functions operate on entire dimensions at once and take advantage of low-level optimizations to perform calculations fast.
For example, the snippet below shows that squaring a numpy array using a vectorized function is >35 times faster than looping over the array with apply_along_axis.
In \[1\]: import numpy as np
In \[2\]: a = np.arange(1000)
In \[3\]: %timeit np.apply_along_axis(lambda x: x\*\*2, 0, a)
62 µs ± 4.23 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
In \[4\]: %timeit a \*\* 2
1\.75 µs ± 38.3 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
2. Don’t forget about I/O¶
Reading and writing data can be an easy bottleneck to overlook, I/O handling often amounts to a few lines of code at the beginning and end of a script. When working with a lot of data—whatever that means to you—carefully consider how data is being stored and accessed. Some tips:
Only load necessary columns into memory and specify data types so the loader doesn’t have to infer them.
Loading data in chunks, processing the chunks, then aggregating is a powerful strategy for efficiently dealing with larger datasets.
Delimited text files aren’t always the right way to store data. Alternative options depending on how your data is structured could include
HDF,NetCDF,Parquet, or a trusty relational database.
3. Be Lazy¶
As thrilling as it may be to write something from scratch, utilizing the right library is a shortcut to writing fast code. Scientific programming libraries are written with speed in mind and take liberal advantage of low-level optimizations (the numpy package is about 50% C code).
4. Think like a computer scientist¶
Micro-optimizations that speed up single expressions are often trumped by picking a more efficient algorithm. Selecting the right data structures and algorithm for the problem at hand is an essential skill when writing efficient code.
Not everyone who writes code for scientific analysis needs a computer science degree and an A+ in an algorithms class. Just like with cars or bikes, a little bit of knowledge of the mechanics of what is going on beneath the hood goes a long way.