tl;dr: We compare Python performance against Clojure, Julia, and Java in text-processing and dynamic collections. Remarkably, Python does well.

Note - this post embeds gists for code. These do not render well on Planet pages or on RSS readers like feedly. It may be better to view on the original site

Situation

I mostly understand numeric performance; I am ignorant when it comes to text and basic data wrangling. Actually, that’s an understatement

The SciPy community has optimized the heck out of numerics; lets think about text!

I think that many of us are in the same position. At work we analyze non-numeric data and run-times now slow down my work cycle. After benchmarking and tuning code within Python I have the following three options to further increase performance:

  1. Use more machines
  2. Invent a better algorithm
  3. Switch to a lower level language

Often we choose number 1, “use more machines”. We make this choice because it’s easy. Option 2, “invent better algorithms” is hard and so we avoid it if we’re not intellectually interested in the problem.

In numeric work we often pursue option 3, “switch to a lower level language” and with good reason. The Python -> numpy -> fancy-numeric-package -> tuned C/Fortran -> CUDA progression often comes along with somewhat predictable order-of-magnitude gains at each step. We know we can improve computation and we know roughly by how much.

In text-based work I lack the same strong intuition on performance that I have in numeric work. I don’t know how valuable lower level languages really are. This blogpost helps to answer that question.

A Small Language Shootout

Can my 30-core shared memory machine or large distributed system be replaced with a few cores running tight compiled code? To test this we run a very simple parse-and-group operation in three languages of current interest:

  • Python – our favorite low-performance language
  • Clojure – A compiled lisp on the JVM with a vibrant community
  • Julia – The language all of Scientific Python is talking about but no one seems to have used.

Each is a modern high-productivity language optimized for development time as well as performance. I would feel comfortable marrying myself to any of them long-term.

Later on in the blogpost I introduce Java as a baseline language:

  • Java – the oddly effective language that everyone loves to hate

Installing Julia

Julia is a compiled language that targets

  • imperative array code (like C/Fortran)
  • with a lightweight syntax (like Python)
  • but with a real type system (like Haskell).

Along with Rust and Go it is one of the recent advances in imperative languages. It caters more to the Matlab/Fortran crowd than the Systems/C crowd (like Go).

While curious about Julia I’ve never played with it. I had been warned about Julia’s installation process. I was told that it depended on a custom LLVM and took up a Gig+ of storage. I was pleasantly surprised when the following worked (Ubuntu 13.04)

mrocklin@notebook:~$ sudo add-apt-repository ppa:staticfloat/juliareleases
mrocklin@notebook:~$ sudo apt-get update
mrocklin@notebook:~$ sudo apt-get install julia
    ...
    Need to get 12.8 MB of archives.
    After this operation, 37.5 MB of additional disk space will be used.
    ...
mrocklin@notebook:~$ julia
               _
   _       _ _(_)_     |  A fresh approach to technical computing
  (_)     | (_) (_)    |  Documentation: http://docs.julialang.org
   _ _   _| |_  __ _   |  Type "help()" to list help topics
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.2.0 (2013-11-16 23:48 UTC)
 _/ |\__'_|_|_|\__'_|  |  Official http://julialang.org release
|__/                   |  x86_64-linux-gnu

julia> 1 + 2
3

Well that was simple! There is no longer an excuse not to try Julia.

Problem

I want to test

  • File I/O
  • Basic string operations
  • Grouping operations (mostly dictionary lookups and collection appends).

To do this we take all of the word-pairs in “Tale of Two Cities” and group them by the first word.

Given data that looks like the the following:

$ cat data.txt
a,b
a,c
a,d
b,a
b,d
d,c
d,a

We produce data that looks like the following:

{'a': ['b', 'c', 'd'],
 'b': ['a', 'd'],
 'd': ['c', 'a']}

But instead of a, b, c we use words from a long book, The Tale of Two Cities

it, was
was, the
the, best
best, of
of, times
times, it
it, was
was, the
the, worst
worst, of
of, times
...

We read in a file, split each line by commas, and then perform a groupby operation. Comma spliting and grouping are both commonly used operations. For example these steps might be used to create a Markov chain to power word prediction in your smartphone keyboard.

Solutions

Note that these are all done using the pure language. Both Python and Julia have a DataFrame project (like pandas) with heavily optimized groupby operations. Today we stick with the core language.

Python

We use toolz for the groupby operation

python benchmark.py word-pairs.txt

Julia

We first make a groupby operation in Julia.

Afterwards out code closely matches the Python Solution

julia benchmark.jl word-pairs.txt

Clojure

The Clojure standard library has everything we need

lein uberjar
java -jar location-of-standalone.jar word-pairs.txt

Numeric Results

The results surprised me.

Python:     200 ms
Julia:      200 to 800 ms  # I don't know what's going on here
Clojure:    550 ms

I expected Python to be dead last. Instead it comfortably hums along in first. It also has the shortest latency (both Julia and Clojure have painful compile times (not included in totals)).

Perhaps this is because neither the Clojure nor Julia solutions have sufficient type information. If any Clojurians or Julians (is that what we call you?) are around I welcome better solutions.

In particular, I was sad to learn that Julia’s readlines function is of type file -> Array{Any} rather than file -> Array{String}. This propagates down to word_pairs being of type Array{Array{Any}} which I suspect stops many meaningful optimizations.

Thoughts on the Code

Timing Macros

Writing code to time other code is tricky. It always requires some sort of super-code. In Clojure and Julia this is the time macro

;; Clojure
(time (... ))

# Julia
@time begin
    ...
end

In Python I used the duration context manager. Context managers commonly serve well where macros are desired.

with duration():
    ...

But normally people use IPython’s timeit magic (also macro-like). I used duration here because it was pure Python and because it doesn’t benefit from the caching of repeated iterations.

Lambdas

I miss nice lambda syntax. Both Clojure and Julia have concise multi-line anonymous functions. Julia even has pretty ones:

Python  : lambda x: x + 1
Clojure : #(+ 1 %)
Julia   : x -> x + 1

As much as I love Clojure I have to say, that’s an ugly lambda. It looks like a child mashing the top row of the keyboard.

No Obligatory Types

There is no necessary explicit type information in any of the languages. This seems to be a common trait among high productivity languages.

Performance from Optional Types?

Clojure and Julia both support adding optional type information to increase performance. Python3 supports static type annotations but doesn’t use them meaningfully. I suspect that one can get more performance on both the Clojure and Julia solutions by adding type information. If any experts are out there on supplying type hints I’d be grateful for the suggestions.

Lets Consider Java

To see how much runtimes could improve we test Java, a language without fanciness, where all types are explicit, and whose compiler optimizations we mostly understand.

The results?

Python:     200 ms
Julia:      200 to 800 ms  # I don't know what's going on here
Clojure:    550 ms

Java:       190 ms

Surprisingly, Java is not significantly faster. This provides us with a pretty stable lower bound for this problem. We’re unlikely to get much faster with standard data structures.

As a corollary this probably answers the question:

“Can more type information accelerate the Clojure/Julia solutions to this problem?”

with the answer:

“No”.

Code

The Java standard library has the data structures we need but fancy operations like groupby are absent and difficult to create.

javac Benchmark.java
java Benchmark word-pairs.txt

Conclusion

It turns out that runtimes for this problem are dominated by data structures for which Python is well optimized. But how common in this computational focus among text-analytics problems? If anyone has insight here I’d love to hear about it. Remember that my goal is to generate intuition about computational performance in text-analytic operations.

Numerical computation intuition tells us that types are important. Switching from a Python int to a C int often results in a significant performance gain. This intuition didn’t carry over well to text and grouping; instead we mostly benchmarked the dictionary/HashMap of each language.

I used to feel guilty about using Python for data processing. But this test gives me confidence in the performance of Python data structures relative to other languages and makes me more optimistic about Python’s role in data analytics in general.

Appendix

My Python duration context manager. This is more a lesson on the value and simplicity of context managers than anything else:

Also, if you want the data then you should run this script:

It depends on toolz which you can get from PyPI with

pip install toolz

or

easy_install toolz


blog comments powered by Disqus