A faster, leaner JavaScript for scientific computing: what features should I keep?

Question

Here I'm really interested in lowering barriers to mathematical education.

Target:

I'd like to see created for the JavaScript community, an equivalent of the Python-based/linked scientific and high-performance computing libraries (great lists of which are available through Sage and otherwise). And I want that, because I'd like to make it easy for people who learn JavaScript to get into scientific and numerical computing without having to learn Python (& company). (I know it's easy to learn Python, as I basically did it at some point, but this suggests that perhaps it'll be easy to compile some restricted subset of JavaScript to Python.)

Hypothesised method:

I'm primarily interested in a new language with minimal difference from JavaScript, because the market ("human compilers") I'm targeting are programmers who already know JavaScript. What I want to target those people for, is to give them a minimally different language in which to write code that compiles to faster C, in the manner that RPython and Cython do for Python. I'm willing to throw out a lot of JavaScript features, I just want to be careful to add a minimum number of features back in. I'll definitely be looking at Lua, Dart, ECMA Harmony (which has no formal date of release, or am I mistaken?), etc. as these are all close resemblances to contemporary (2012) implementations of JavaScript.

Questionable Motivations:

I'm personally willing to learn any language/toolset that gets things done faster (I'm learning Erlang myself, for this), but here, I am specifically interested in lowering the bar (sorry) for other people who may not have such willingness. This is just one of those "want to have my cake, and eat it too, so I am putting some time into researching the problem" situations. I have very limited prior experience in computer language design, but so far from a hacking-the-ecosystem point of view, the problem seems interesting enough to study, so, I hope to be doing more of that soon.

Answer 1

If you want to use JavaScript for scientific computing, why not try Node.js? Node is based on Google's V8 engine with some additional functionality for interacting with the file system, writing servers and linking to C and C++ libraries (this is definitely not an exhaustive list). It sports fast execution times and provides an interactive console as well as a way to execute your programs from the command line. I found @DaveClarke's comment about Python acting as "glue code" interesting because of the wording. Here is an excerpt from the Node documentation on addons:

Addons are dynamically linked shared objects. They can provide glue to C and C++ libraries.

Node also has an excellent package manager called NPM, which would make it easy to share scientific modules which you have written. The NPM repository already contains some scientific modules, such as "natural" (for natural language processing) or "clusterfck" (a clustering library). There are most definitely many other libraries out there for a variety of applications.

The great thing about this solution is that you wouldn't have to modify JavaScript or create a new language. The syntax is essentially the same, but your code does not need a browser to execute. If you plan to follow up with node, a good first step might be to start creating addons for existing open-source scientific libraries written in C and C++, or you could port them to JavaScript. I was actually thinking about using node for scientific programming today which is what led me to this question. It would be great to see more robust tools like you find in Python.

I initially had links to natural and clusterfck, but I do not have enough rep to post them. Luckily, they come up as the first results if you google "clusterfck node" and "natural node".

Answer 2

The point of the Python scientific libraries is that Python merely acts as a glue code for shuffling around the data between the scientific libraries. Python itself does little of the high performance grind work. Python is a good language for this, as it supports both the procedural and object-oriented paradigms (and functional to a large degree), and one can easily link with external libraries.

Javascript is heavily tied to browser technologies. It is mostly an awful language not at all designed to be efficient or to write large scale software, although in practice both of those happen due to huge investments in compiler technology and careful use of the language, respectively. Javascript is made 1000 times nicer by well designed libraries such as JQuery. I'm not sure whether Javascipt provides any support for linking easily with C libraries (though I guess it must, indirectly through the browser, and via Java).

I don't mean to be discouraging, but the whole idea seems ill-founded. Javascript and high-performance computing are only found together in sentences beginning with "A man walked into a bar ...."

That said, I'm willing to be proven wrong. Real-time Java exists, which is an odd combination, and is reasonable successful, so who knows.

Answer 3

Just an idea: you can use scientific libraries written in Java (I think you'll find a lot of work already available) with Javascript - that has a syntax similar to Java - used only as a wrapping language for:

• scripting
• user interaction
• (re)modeling and standardization of the underlying high level classes/packages/functions ... though I think it would be better to do this in the Java "layer"

See the Rhino project for a stable Javascript interpreter written in Java.

If you want you can use the Java JNI to use scientific libraries written in C/C++.

The whole picture:

                              ____ disparate Java libraries
Javascript   ----- Java      /
(scripting,      ("package   \____ JNI ____ C/C++ libraries
 interaction,     remodeling")   
high-level stuff)

EDIT: just Googling around a found stuff like this: Java GNU Scientific library, List of (Java) numerical libraries, Java Graph library, jHepWork Multiplatform environment for scientific computation and data analysis

Answer 4

Creating a new language is not necessary. One can write performant numeric computing in JavaScript today, without compilation/transpilation. As a scripting language, JavaScript compares quite favorably to Python and R. For benchmarks, run the language benchmarks in stdlib, which compares performance across a wide-range of APIs.

For server-side applications, one can write Node.js native add-ons to interface with the same C/C++/Fortran libraries used in other numeric computing environments, such as R, Python, and Julia. As shown in my work, Node.js native add-ons allow linking to hardware optimized numeric computation libraries, thus enabling high-performance numeric computation from JavaScript.

For client-side applications, one can compile to WebAssembly to achieve 1.5-2x of native performance. For older environments, one can compile to asm.js, which is a reduced subset of JavaScript more amenable to compiler optimization. In my work, asm.js can match naive reference (i.e., non-hardware optimized) implementations written in C/C++/Fortran.

In short, there are no technical reasons why you cannot write numeric and scientific computing libraries in JavaScript, itself. What is needed are people with the time and wherewithal to write rigorous and robust implementations.

disclosure: I and others are working on the stdlib project, which aims to bring robust, high-performance numeric computing capabilities to JavaScript and Node.js.