Why is bcryptjs slower on AWS Lambda than in local docker?

Question

I have Lambda written in NodeJS. I noticed it takes several seconds to complete. I added logs and found that bcrypt is quite slow.

Packages:

"dependencies": {
  "bcryptjs": "^2.4.3",

Source code:

const bcrypt = require('bcryptjs');

console.log("User was found"); // following part takes more than 1 second!
if (bcrypt.compareSync(password, user.password)) {
    console.log("Password verified"); //

This is a log from AWS LogWatch:

2020-01-13T20:25:30.951 User was found
2020-01-13T20:25:32.670 Password verified

and

2020-01-13T20:31:20.192 User was found
2020-01-13T20:31:21.550 Password verified

So it takes 1.7 seconds. I ran the same code in docker on my machine

2020-01-13T20:09:48.109 User was found
2020-01-13T20:09:48.211 Password verified

Locally it takes just 120 ms. AWS uses NodeJS 10.x, local docker image is probably 8.x. I do not know how to tell docker to reflect changes in packaged.yaml.

Is this NodeJS regression? Or some issue on AWS configuration?

Answer 1

Encryption performance is typically CPU bound. AWS Lambda CPU is proportional to RAM, so you should choose the largest (3008 MB) and re-test.

When I run this inside a Lambda function handler on a 3008 MB RAM Lambda in us-east-1, the compareSync call consistently takes 90-100ms. With a 128 MB Lambda, it takes a little over 1s.

On a related note, it's helpful to understand that choosing the lowest (128 MB) RAM option, simply because it is cheaper per GB-s, is not always the best thing to do. While the highest RAM option (with proportionally higher CPU and network) is certainly more expensive per GB-s, it also completes Lambda functions a lot quicker. So, for example, you might be able to complete your task in 1/10th of the time at only 1.75x the cost. In a lot of situations, that can be very valuable.

There is a project that can help you tune price/performance for your Lambdas: alexcasalboni/aws-lambda-power-tuning