Overview of Commercial Workload Tuning

Using a commercial workload to benchmark a Linux server is the best way to fully cover most, if not all, subsystems in the server. The application server, which is the typical mode for servicing commercial applications, involves many layers of the software performance stack described in Chapter 21, "Case Study: Network Performance on Linux."

Performance analysis of an application server is a system-wide process that involves analyzing the network configuration, the file system, the JVM, the kernel scheduler, memory usage, and CPU utilization. Performance analysis might also require analyzing the operating system and obtaining kernel traces and kernel profiling information, as well as obtaining information from the JVM itself using the JVM's debugging and tracing utilities. Last, performance analysis might include gathering performance data on the application itself with an application-level profiler.

Instead of presenting an actual commercial workload of a specific enterprise, this chapter works with commercial workload models, which are typically used in performance benchmarks and analysis. Commercial workload models are also often used in reference implementations of new programming model specifications. For example, when J2EE was introduced, a sample application called PetStore was used to illustrate some of the best practices for J2EE application development. PetStore was also intended for customers to use when building their own enterprise web applications. PetStore is a commercial workload model for online retail applications. Sun maintains a blueprint of the Java PetStore at http://java.sun.com/blueprints/.

Commercial workloads are typically simulated using a representative application for a given application domain. The performance of an application server can then be determined by tuning the application server and the system platform. Most performance testing of commercial workloads is done by gradually increasing the load against the server until a saturation point is reached.

One of the hardest parts of performance tuning is making sense of the performance data that has been collected and determining the problem that needs analysis.