Skip to content

Latest commit

 

History

History
268 lines (198 loc) · 11.7 KB

nodejs_bindings.md

File metadata and controls

268 lines (198 loc) · 11.7 KB

Making OSRM Matrix Computations with its NodeJS Bindings

In this tutorial we would like to demonstrate the power of Open Source Routing Machine utilizing its NodeJS bindings without any overhead of running a HTTP service. If it's the first time you have heard of bindings, just imagine them as an application programming interface (API) that provides glue code specifically made to allow a programming language to use a foreign library which is not native to that language. In this scenario our foreign library is OSRM written in C++ which is being consumed by JavaScript. Additionally, we would like to give you a small taste of how you could optimize for performance when requesting large square matrices.

If you want to take a glimpse at the code straight away, you can get the it here.

For a comprehensive feature list and a more general overview of FOSS routing engines, read our article.

Introduction

Open Source Routing Machine is a high performance open source routing software (BSD-2-Clause License) written in C++ and mainly designed to consume OpenStreetMap data. The logic required to run OSRM services directly from NodeJS is ensured by exposing its compiled binaries, e.g. you don't have to separately install and build OSRM on your machine.

While we are interested in the matrix (table) capabilities, you also have the option to use its other services, too.

Disclaimer
Validity only confirmed for Ubuntu 20.04 / Mac OSX 10.15.6 and OSRM 5.26.0.

Recommended

  • A basic understanding of Javascript, NodeJS and npm
  • The awareness that a red printed terminal line doesn't necessarily mean, it's an error ;)...

This tutorial will guide you through the process of how to configure and run the NodeJS bindings for OSRM.

So let's go, open your terminal!

1. Prerequisites

Test your NodeJS & npm installation

Test whether your environment already has NodeJS installed.

node -v

Similarly test if your environment has Node's package manager npm installed.

npm -v

If either output is similar to this:

command not found: ...

You will have to install nodejs and/or npm first. To do this have a look at the How To Install Node.js on Ubuntu 20.04 or if you are running OSX head over to Install Node.js and npm using Homebrew on OSX and macOS.

Create your Working Directory and get NPM ready

First create a working directory:

mkdir ~/gisops_osrm_nodejs && cd $_

In order to install the required dependencies for this tutorial you will need to copy and paste the following snippet into your package.json file.

{
  "name": "osrm node bindings gis-ops tutorial",
  "version": "0.1",
  "dependencies": {
    "async": "^3.2.2",
    "osrm": "^5.26.0"
  }
}

We are making use of 2 libraries osrm and async which will allow us to use the OSRM nodejs bindings in a multithreaded fashion. With the following command you instruct npm to install these dependencies which will read the package.json file sitting in the same folder.

npm install

Note
OSRM bindings binaries are currently only available for the last three LTS releases and the current release of nodejs, so make sure you're using a compatible version (e.g. by using nvm). Alternatively, you can build from source.

If you check the contents of your folder you will notice a new folder named node_modules. In there you'll find both of these libraries including their dependencies.

2. Prepare the OSRM Graph

Extract & Contract your OSM File

For this demo, we are going to use the OpenStreetMap data of Berlin which you can simply download into your working directory.

wget https://download.geofabrik.de/europe/germany/berlin-latest.osm.pbf

Afterwards we have to extract the OSM data which is required by OSRM to generate the topology and then contract it which boosts the performance. More can be read here. We are using the OSM extract which we just downloaded and using the default car.lua file residing in the dependency folder. You obviously have the option to change to a different profile or tailor the lua file for your needs. Some examples can be found here.

a. node_modules/osrm/lib/binding/osrm-extract berlin-latest.osm.pbf -p node_modules/osrm/profiles/car.lua

b. node_modules/osrm/lib/binding/osrm-contract berlin-latest.osrm

3. Compute Matrices

Create a Single OSRM Matrix Request

Moving forward we will create a small script which will use the NodeJS bindings and prepare a matrix request. The OSRM method is the main constructor for creating an OSRM instance which requires a .osrm dataset. Afterwards we can construct a request which will be sent to the the osrm.table function.
For the sake of this example we are using a square matrix of 1.000 random coordinates in and around Berlin. We ask OSRM for durations and distances and expect a many-to-many response.

// the nodejs bindings
const OSRM = require("osrm");

// importing random coordinates in Berlin used as sources and destinations
const randomCoordinates = require("./random_berlin_coordinates");
const coordinates = randomCoordinates.coordinates1000;

// teaching the bindings to use the osrm graph prepared in the previous step
const osrm = new OSRM("berlin-latest.osrm");

// a small function to create the osrm options
// https://github.com/Project-OSRM/osrm-backend/blob/master/docs/nodejs/api.md
const makeOsrmOptions = (sources, destinations) => {
  return {
    coordinates: coordinates,
    sources: sources || [],
    destinations: destinations || [],
    annotations: ["distance", "duration"]
  }
}

console.time("osrmTableSingleRequest");
const osrmOptions = makeOsrmOptions()
osrm.table(osrmOptions, (err, result) => {
  if (err) {
    console.log(err)
  }
  // console.log(result.durations, result.distances)
  console.timeEnd("osrmTableSingleRequest");
});

Depending on the hardware computing this request it will take approximately 1-1.5 seconds to return the matrix result of 1.000.000 pairs.

Optimizing your Single Matrix Request

If you want to be a little more adventurous, you can break your large matrix request into bins which will yield a little speed up. We can do this with some handy utility functions which we will add to our script moving forward. The chunkSize variable determines how large your sub-matrices should be - we call them bins. These are created by the makeBins function which does nothing other than loops over your many-to-many matrix and sticks ranges into bins. With our settings we are creating exactly 4 bins of many-to-many matrices, i.e. [0-500] to [0-500], [0-500] to [501-1000], [501-1000] to [0-500] and [501-1000] to [501-1000].

const chunkSize = 500

const range = (start, end) => {
  return Array(end - start + 1).fill().map((_, idx) => start + idx)
}

const makeBins = (coordinates, chunkSize) => {
  const bins = []
  for (let x=0; x < coordinates.length; x += chunkSize) {
    let endX = x + chunkSize - 1;
    if (endX > coordinates.length - 1) {
      endX = coordinates.length - 1;
    }
    for (let y=0; y < coordinates.length; y += chunkSize) {
        let endY = y + chunkSize - 1;
        if (endY > coordinates.length - 1) {
          endY = coordinates.length - 1;
        }
        bins.push({ sources: range(x, endX), destinations: range(y, endY) })
    }
  }
  return bins;
}

const bins = makeBins(coordinates, chunkSize)

At this point we are ready to send off our bins of coordinates. We could do this sequentially which wouldn't yield much benefit, however we could also make heavy use of OSRM's powerful architecture of distributing requests across multiple cores. By default it uses 4 but you have the option to give it more to work with. There exists a handy NodeJS library called async which will help us achieve what we need, namely sending multiple requests to OSRM at once, waiting for all computations to finish and ultimately stitching them back together again. Specifically we are using one of its core functions mapLimit to do this. We give it all of our bins and a limit parameter (which we call parallelism because it sounds more fancy) governing the maximum number of async operations at a time. Once all callbacks have been successfully invoked, we can optionally loop over all results and stitch them back together again.

/** 
... existing header
 */

const async = require("async");

/** 
... existing code 
 */


// as the name suggestes, it stitches the results back together!
const stitchResults = (results) => {
  // make empty lists which will be filled up
  const durations = Array.from({ length: coordinates.length }).fill().map(item => ([]))
  const distances = Array.from({ length: coordinates.length }).fill().map(item => ([]))
  for (const matrix of results) {
    const minSourceIdx = Math.min(...matrix.sources)
    const minDestIdx = Math.min(...matrix.destinations)
    for (const sourceIdx of matrix.sources) {
      for (const destinationIdx of matrix.destinations ) {
        durations[sourceIdx][destinationIdx] = matrix.durations[sourceIdx - minSourceIdx][destinationIdx - minDestIdx];
        distances[sourceIdx][destinationIdx] = matrix.distances[sourceIdx - minSourceIdx][destinationIdx - minDestIdx];
      }
    }
  }
  return { durations, distances }  
}


// how many requests to have in flight at once
const parallelism = 4

console.time("osrmTableParallel");
async.mapLimit(
  bins,
  parallelism,
  (bin, callback) => {
    // pass in coordinates again but limit osrm.table to compute only sources and destinations of the bin
    const osrmOptions = makeOsrmOptions(coordinates, bin.sources, bin.destinations)

    osrm.table(osrmOptions, (err, result) => {
      if (err) {
        callback(null, result);
      }
      callback(null, { ...result, sources: bin.sources, destinations: bin.destinations } );
    });
  },
  (err, results) => {
    if (err) throw err;
    // if you want to stitch your OSRM computations back together again
    const fullMatrix = stitchResults(results)
    console.timeEnd("osrmTableParallel");
  }
);

Depending on your hardware, you will notice that this approach averages around 700-900ms, so approximately 30-50% faster than the single 1.000 x 1.000 request which we started off with.