ASEP BAGJA

Recently, I’m planning to build a weekend project for personal purposes. It’s an application to analyse stock market data for helping me in my investment journey. I plan to use Node.js as the backend. However, I’m curious about how Node.js perform when handling a large dataset. How fast can it do the operation?

To answer that question, I created this scenario:

1. Create an array with 5 million elements with a float data type. It is to simulate the stock price dataset.
2. Calculate the Simple Moving Average for a specific window interval from that dataset. It is the operation that I want to measure.

Here is how I generated the 5 million elements of an array. We won’t measure this operation.

const simulateClosingPrices = Array.from({length: 5000000}, () => Math.random() * 5000);

Below is the code to calculate SMA in JavaScript. I packaged it as a Node module.

Here is the execution time.

elapsed time:   160.32ms
elapsed user:   124.93ms
elapsed system: 28.68ms

Can we make it faster?

Hmm…how if I involve Rust? I love that language. It’s a system programming language that doesn’t have a garbage collector and compile into native code. In theory, it will be faster than JavaScript. To embed and use Rust with Node.js, I have to package it as a Node module. There are two options to do that. First, compile it as a Node native module. The second one, target it for WebAssembly and package it as a Node module.

Let’s try both options. These are the toolchain to make the Node module with Rust:

• wasm-pack: It’s a toolchain to compile Rust code into WebAssembly.
• node-bindgen: It’s a toolchain to compile Rust code into a native module.
• neon-bindings: This one can compile Rust code into a native module too. But, after I tried it, I couldn’t reuse my existing Rust code for WebAssembly. I must adjust many parts of it. So, I abandon it.

Below is the code to calculate SMA in Rust and targeting WebAssembly.

Below is the code to calculate SMA in Rust and targeting the native module. In my case, it will compile to aarch64-apple-darwin instruction since I ran this benchmark on the Apple M1 processor.

The result

First run
|                           | JS     | WASM  | Native Module |
|---------------------------|--------|-------|---------------|
| elapsed total time in ms  | 160.32 | 81.89 | 1137.70       |
| elapsed user time in ms   | 124.93 | 53.81 | 1106.9        |
| elapsed system time in ms | 28.68  | 22.25 | 74.05         |

Second run
|                           | JS     | WASM  | Native Module |
|---------------------------|--------|-------|---------------|
| elapsed total time in ms  | 144.09 | 75.94 | 1165.27       |
| elapsed user time in ms   | 123.32 | 52.72 | 1109.09       |
| elapsed system time in ms | 22.53  | 23.26 | 84.15         |

It is surprised me. My expectation was the native module will perform faster than WebAssembly, but the result was the opposite. Probably, it happened because I tried to map the JS array directly into the Rust vector. I read some articles it said would be better if I passed the TypedArray instead of JS array. Let’s try it in the next article.

According to this experiment, WebAssembly performs 2x faster for the total time than pure JavaScript when handling this scenario.

The featured image photo by Chris Liverani on Unsplash