Performance Testing Serverside Applications

Umer Mansoor Umer Mansoor Follow Nov 16, 2016 · 3 mins read

Performance testing server-side applications is a crucial process to help understand how the application behaves under load. It helps software teams fine-tune their applications to get the best performance while keeping the infrastructure costs low. Performance testing answers several important questions such as:

  • Is the application ready to handle the traffic that’s going to hit it?
  • What do average response times and latencies look like under normal and peak loads?
  • Can the application be scaled out?
  • What are the bottlenecks? (could be CPU, memory, an external service or a database server)
  • How many instances are needed for supporting the estimated traffic (i.e. max RPS)?
  • What type of instances are needed? Does the application requires an instance with higher CPU to Memory ratio? Or does it need an instance type that supports high network utilization?
  • Does the application slowly degrade in performance under load? Is it slowly leaking a resource that eventually crashes it after a few hours or days?

Leaky Pipe How much load can your application sustain before it buckles under?

A lesson I learned recently is that performance testing should not be an after-thought. Software teams should start performance testing early in the release cycle and not wait until the end to do it. I once worked on a team that built a backend service that passed all unit, integration and end-to-end tests with flying colors. QA engineers didn’t find any bugs in the application’s logic. However, the performance was just terrible when we ran load tests on it. On a single m4.large instance, the application supported 80% fewer requests than the team had estimated! The main bottleneck was found to be the 2-core CPU that was utilized to its maximum capacity as the application issued several queries to the database and applied complex algorithms to build a graph. Investigations by developers revealed that to reduce the amount of work the CPU was doing, it would require significant design changes. But it was already too late - the deadline was just weeks away. We decided to proceed with the release - albeit by over-provisioning the hardware and over-running our cost estimates by a factor of 3.

Performance testing is a broad topic. Teams I work with run load and soak tests to measure performance metrics such as throughput, latency, resource utilization, etc. using a wide variety of tools. At Glu, we build REST services in Java and use the following tools for our performance tests:

  • YourKit Profiler to profile CPU and memory usage at a fine-grained level.
  • Apache Jmeter to generate load (in reality, Blazemeter or distributed Jmeter)
  • Amazon Cloudwatch to monitor resource utilization because we deploy our services on the Amazon cloud.
  • Hosted Graphite to observe custom metrics that the application generates and we are interested in.
  • Kibana dashboards to look at the logs, errors, etc.

Before I wrap this post up, there are few other important lessons I’d like to share:

  • Run performance tests on a production-like environment. I have seen teams run performance tests on their MacBooks Pros with 8-core CPU’s and get drastically different results than from the actual cloud instances with puny, virtualized hardware.
  • Create a good load test plan which requires careful thought. The goal is to emulate the load that real users would generate otherwise you might spend a lot of time chasing ghosts and fixing issues that are unlikely to happen in production. For example, I was investigating an issue with the performance of a Chat service under load. After looking at the load test script, I found that it grouped tens of thousands of users together. In reality, a group has on average about 10-50 people. The fix was to update the load test script to use a random group-id or a group-id from a pool instead of reusing the same id for each request.

I hope this post was helpful. Would love to hear your thoughts, the tools and the approach you take for performance testing. Till next time.



Umer Mansoor
Written by Umer Mansoor Follow
Hi, I am Umer. I live in the San Francisco Bay Area and it is where I work. Since 2011, I have worked at startups, mobile gaming and internet companies. If you have any questions, feel free to message me.

Speak Your Mind

Join Newsletter
Enter your email address to receive a very occasional newsletter. I'll never spam or sell your email.