16 Performance optimization

28 Apr 2016

First of all, scientific code must provide correct answers: you need to know the level of your approximations, i.e., you need to know the errors in your computations, and you need to be able to test your code for correctness.

But secondly, you typically also want to run your code efficiently and as fast as possible. For this, you want to optimize code.

However, as Donald Knuth said ¹: Premature optimization is the root of all evil. What this means is that you should first write correct code and then carefully evaluate where the performance bottlenecks in your code are. Only then spend time optimizing code (while continuously testing to check that you are not introducing new bugs).

In this lesson you will learn to use a tool named a profiler to measure the performance of pieces of your code. The Python profiler will then tell you where you should spend your optimization efforts. Examples will be given how to boost the performance of the MD code from the Midterm Project. This includes moving expensive calculations out of inner loops and using numpy expressions instead of loops.

---

References