Java vs. Go Microservices — Load Test With Multiple Users

Microservice architecture allows us to choose between technologies and programming languages when creating new application services. What language should we choose to serve more users on the same hardware? To answer this question, it will be good to know performance differences between the same applications written in Java and Go.

Prerequisites

• No additional performance enhancements should be applied. Use minimum configurations with default frameworks and libraries settings.

• No ORM frameworks. Use pure DB drivers and the same SQL queries. Postgres JDBC 4.2 driver for Java and github.com/lib/pq for Go.

How to

1. Create simple Java/Go REST API applications with DB (Postgres) data storage.

2. Create load tests with JMeter or a similar tool.

3. Run the applications, load tests, and database on separate AWS instances.

4. Collect the load test results.

Systems Under Test

As systems under test, I've prepared two bank applications: bank-java and bank-go.

This is probably the simplest bank application in the world. The only things it can do are create new clients and transfer money from one client to another.

APIs:

• POST /client/new/{balance}  - create new client with initial balance.

• POST /transaction  - moves money from one account to another.

• GET /client/{id}/balance - returns current balance for client.

Frameworks and Dependencies

When choosing frameworks and libraries, I used the most recent, popular, and easy ones to get the application ready as quickly as possible.

Bank-java: Java 10, Spring Boot 2.0.4, spring-web 5.0.8, PostgreSQL JDBC 4.2.4.

Bank-go: Go 1.8, gorilla/mux, github.com/lib/pq.

Bank Applications Source Code

Bank-java: https://github.com/nikitsenka/bank-java

Bank-go: https://github.com/nikitsenka/bank-go

Test Project

The test project, Bank-test, performs calls to the bank APIs with a dynamically changing number of users, from 1,000 to 10,000, verifies responses, and collect statistics.

Test Environment

AWS was chosen as the test environment. Next, AWS EC2 instances were created:

1. Bank-go t2.micro (Variable ECUs, 1 vCPUs, 2.5 GHz, Intel Xeon Family, 1 GiB memory, EBS only)

2. Bank-java t2.micro (Variable ECUs, 1 vCPUs, 2.5 GHz, Intel Xeon Family, 1 GiB memory, EBS only)

3. Postgres d2.xlarge (14 ECUs, 4 vCPUs, 2.4 GHz, Intel Xeon E52676v3, 30.5 GiB memory, 3 x 2048 GiB Storage Capacity)

4. Bank-test t2.2xlarge (Variable ECUs, 8 vCPUs, 2.3 GHz, Intel Broadwell E5-2686v4, 32 GiB memory, EBS only)

👁 Image

Results

The full results log can be found here.

Results Aggregation

👁 Chart

👁 Chart

Results Summary

Both applications work perfectly with 1,000 simultaneous users. With 2,000 users, Go performance reduces significantly, whereas Java remains perfect. Starting from 3,000 users and above, both applications show an unacceptable response time, and the number of error responses grows significantly.

Conclusion

Using the same hardware, the Java REST API application can serve twice as many simultaneous users as the Go application with a PostgreSQL database.