Introduction and Motivation

Objectives

• Know what to expect from this course

• Build a general, programming-language agnostic notion of performance

Instructor note

• Inform the format of this talk

  • Mix of talk, type-along and demos

• 5 min teaching

Python and its defacto implementation CPython is now widely used for a spectrum of applications. It has now experienced practitioners doing web-development, analytics, research and data science. This is possible because of the following traits of the Python ecosystem:

• Batteries included

• High-level programming that abstracts away the technical details

• Mature well-maintained libraries which form a firm foundation, the scientific Python ecosystem, which includes:

and many more…

Extensions: a technical detail hidden in plain sight

A common theme behind the Python standard library and its popular packages is that some parts of the code which are computationally intensive are actually modules or functions which are either:

• interfaced extensions with an external implementation in C, C++, Fortran, Rust…

• source-to-source extensions written in Python or Python-like code, which is compiled ahead-of-time or just-in-time

Extensions can be imported as normal Python functions or modules. There are many tools which help you in creating extensions:

Figure: The coloured bubbles represent programming languages. An outward arrow represents exporting a code into an extension using a runtime or a library. An inward arrow represents importing an extension (or linking to a code using an API, such as Python’s C-API or via a foriegn function interface (FFI). The choices are many!

Discussion

What are the advantages and disadvantages of using Python code written using multiple programming languages under the hood, in terms of software development and maintenance?

Solution

Pros 👍: This approach enables us to:

• build high-level, performant applications which tend to be readable

• focus on the problem at hand, without getting sidetracked with implementation details

• rapidly prototype the experimental parts of the code

• interfacing allows re-use of established codes

Cons 👎: Some known downsides are:

• interfaced codes require knowledge of multiple languages

• compiled codes are harder to debug and rapidly-prototype

Now we have a notion of how extensions work. We could use extensions to address performance issues. However building an extension is quite often the last resort. More on that will be discussed in the next episode.

Different kinds of performance bottlenecks

• I/O bound: the code idles often and is waiting for a disk or network read/write operation to finish. Such bottlenecks can be often remedied by caching, multi-threading or async-programming.

• Memory bound: the data to be processed does not fit in the RAM and the code needs to process data in batches instead. This is often a hardware limitation.

• CPU bound: the code consumes a lot of CPU cycles, often seen by monitoring the system showing 100% CPU usage in 1 core for serial applications, or in all cores for parallel applications. This will be the focus of this course.

Gems of wisdom

Before we dive further into the course it is important to remember some idioms, which is true in the case of most real-world applications.

---

Keypoints

• Find a balance between runtime efficency and cost of development.

• Tests can help in maintain correctness before you change the code.

• CPU-bound or I/O-bound or memory bound?

• Do not optimize everything.

• Creating extensions are one way of improving performance

---

[2]

Although Amdahl’s law is about speedup due to parallelization, we can still associate it with speedup of serial programs. This is because the law is formulated in terms of execution-time.