Serverless Performance with Headless Chrome

The number of people running interesting binaries in AWS Lambda is increasing, and with the recent release of Headless Chromium—a version of the Chrome browser that can run without a UI—there have been experiments and different open-source projects of individuals running it inside an AWS Lambda environment. Running a large, memory-intensive piece of software designed for desktops inside a serverless function presents interesting performance challenges, including how to minimize cold start time and how to optimize the right amount of memory for the function.

This talk is about optimizing performance of chromium inside Lambda using multiple data sources, including AWS X-Ray traces and custom metrics. While the talk is focused on headless chrome in AWS Lambda, the performance techniques can be applied to any memory or compute-intensive serverless functions.

Transcript

1. SERVERLESS
   PERFORMANCE LESSONS
   WITH HEADLESS CHOME
   @smithclay
   ServerlessConf NY, Oct 2017
   

   View Slide

2. SERVERLESS PERF STRATEGY
   ƛ ƛ is it fast?
   ƛ ƛ is it fast concurrently?
   ƛ ƛ what's the worst
   case? why?
   

   View Slide

3. AWS Lambda Function
   index.handler (node.js runtime)
   ./bin/chromedriver
   process.exec
   https://medium.com/clog/running-selenium-and-headless-chrome-on-aws-lambda-fb350458e4df
   ./bin/headless_chrome
   BIG BROWSER, LITTLE
   FUNCTION
   + Simple
   Selenium Script
   

   View Slide

4. AWS LAMBDA: BATTERIES
   INCLUDED DATA SOURCES
   Amazon Cloudwatch Logs
   Amazon Cloudwatch Metrics
   AWS X-Ray Traces
   

   View Slide

5. DISTRIBUTED TRACE
   VISUALIZTION
   1200ms
   400ms
   370ms
   Service A
   Service B
   Service C
   Time
   

   View Slide

6. HELLO, COLD START
   AWS X-Ray Trace: AWS Function invoked from the Developer Console
   

   View Slide

7. COLD VS WARM START
   Notice: Function gets
   called almost
   immediately after the
   service receives the
   trigger.
   

   View Slide

8. COLD STARTS: MORE
   ANALYSIS NEEDED
   Aggregate function
   latencies from X-Ray to
   understand trends
   

   View Slide

9. VISUALIATING TRACES IN
   AGGREGATE
   1) How long does it
   take for my function to
   initialize?
   2) Where is time
   generally being spent?
   

   View Slide

10. FOR THIS FUNCTION, COLD
    STARTS ARE RARE
    Over a 24 hour period,
    the function rarely
    experiences cold
    starts.
    

    View Slide

11. AWS LAMBDA: COMMON
    SETTINGS TO MODIFY
    Language Runtime
    "Memory" Dropdown
    Function size (mb)
    

    View Slide

12. DURATION: BEFORE AND
    AFTER INCREASING MEMORY
    768mb
    1152mb
    Chromium needed more memory!
    

    View Slide

13. WRAP UP
    • Collect data for several days for many function
    runs.
    • Consider enabling x-ray traces for non-trivial
    functions
    • Test scenarios where your function has many
    concurrent executions (i.e. large spike in the
    number of requests)
    • Don't be afraid to increase memory: it might end
    up costing less.
    •Measure and share!
    

    View Slide

14. THANKS!
    slides: @smithclay + speakerdeck.com
    Original blog post: https://blog.newrelic.com/2017/06/20/lambda-
    functions-xray-traces-custom-serverless-metrics/
    

    View Slide