Performance Comparison of Seven WebSocket Server Frameworks and an Additional Go Implementation

Recently I implemented prototypes of WebSocket push servers using several frameworks and conducted a benchmark of seven implementations, later adding an eighth one written in Go.

The seven frameworks tested are:

Netty (http://netty.io/)

Undertow (http://undertow.io/)

Jetty (http://www.eclipse.org/jetty/)

Vert.x (http://vertx.io)

Grizzly (https://grizzly.java.net/)

spray-websocket (https://github.com/wandoulabs/spray-websocket)

nodejs-websocket/Node.js (https://github.com/sitegui/nodejs-websocket)

An additional Go implementation was also tested and performed well.

Test Environment

Three AWS C3.4xlarge servers were used: one as the server and two as client machines. Each client machine launched 10 clients, for a total of 20 clients.

The C3.4xlarge specifications are shown in the image below:

Both server and client machines were tuned following the optimizations described in a previous article.

Test configuration:

20 clients

Setup rate = 500 * 20 requests/second = 10,000 requests/second

Each client creates 50,000 WebSocket connections (total 1,000,000 connections)

After all connections are established, a timestamp message is sent to every client, which measures latency.

If the server setup rate becomes too slow, the test is stopped.

Three performance phases are monitored: during setup, after setup while the application is idle, and during message broadcast.

Test Results

Netty

Setup phase:

CPU idle: 90%

Minor GC: few

Full GC: none

After setup (idle):

CPU idle: 100%

Memory usage: 1.68 GB

Server free memory: 16.3 GB

Message broadcast:

CPU idle: 75%

Minor GC: few

Full GC: none

Message latency (one client):

Vert.x

Setup CPU idle: 95%

Minor GC: few

Full GC: none

Idle CPU idle: 100%

Memory usage: 6.37 GB

Server free memory: 16.3 GB

Message broadcast CPU idle: 47% ~ 76%

Minor GC: few

Full GC: few

Message latency (one client):

Undertow

Setup CPU idle: 90%

Minor GC: few

Full GC: none

Idle CPU idle: 100%

Memory usage: 4.02 GB

Server free memory: 14.2 GB

Message broadcast CPU idle: 65%

Minor GC: few

Full GC: none

Message latency:

Jetty

Setup CPU idle: 2%

Minor GC: many

Full GC: none

Memory usage: 5 GB

Server free memory: 17.2 GB

When about 360,000 connections were being established, setup became very slow, GC frequency increased, and the test was aborted.

Grizzly

Setup CPU idle: 20%

Minor GC: some

Full GC: some

Memory usage: 11.5 GB

Server free memory: 12.3 GB

At roughly 500,000 connections, setup slowed dramatically and frequent GC forced the test to stop.

spray-websocket

Setup CPU idle: 80%

Minor GC: many

Full GC: none

Similarly, around 500,000 connections caused severe slowdown and GC pressure, leading to test termination.

Node.js

Setup CPU idle: 94%

Idle CPU idle: 100%

Memory usage: 5.0 GB

Server free memory: 16.3 GB

Message broadcast CPU idle: 94%

Message latency (one client):

Go

Setup CPU idle: 94%

Idle CPU idle: 100%

Memory usage: 15 GB

Server free memory: 6 GB

Message broadcast CPU idle: 94%

Message latency (one client):

Test Result Analysis

Netty, Go, Node.js, Undertow, and Vert.x can all establish one million connections; Jetty, Grizzly, and Spray cannot.

Netty shows the best performance with very low memory consumption and modest CPU usage.

Jetty, Grizzly, and Spray generate many intermediate objects, causing frequent garbage collection; Jetty performs the worst.

Node.js performs excellently, especially with a single‑instance single‑threaded setup, achieving fast connection speed, low latency, and reasonable memory usage.

Undertow also performs well, with slightly higher memory usage than Netty but comparable overall metrics.

Spray exhibits high CPU usage (≈40 %) even when no messages are sent under heavy connection load.

Source: Bird's Nest

Copyright statement: Content originates from the internet; copyright belongs to the original author. We strive to credit authors and sources unless verification is impossible. If any infringement is identified, please inform us for immediate removal and apology. Thank you.

-END-