Performance Evaluation of Java Stream API: Serial vs Parallel Execution

The article investigates the performance of Java Stream API, questioning whether its convenient syntax incurs a noticeable overhead compared to traditional loops.

All tests run the JVM in -server mode on a CentOS 6.7 server equipped with an Intel Xeon X5675 (6 cores, 12 threads), 96 GB RAM, and JDK 1.8.0_91. To reduce variability, the CMS garbage collector is forced with -XX:+UseConcMarkSweepGC -Xms10G -Xmx10G and JIT compilation is triggered early using -XX:CompileThreshold=10000. Parallel streams use the shared ForkJoinPool.commonPool(), with CPU affinity controlled via the Linux taskset command.

Experiment 1 – Primitive Iteration : Finding the minimum value in an int array. Results show serial streams are roughly twice as slow as a for‑loop, while parallel streams on all 12 cores outperform both.

Experiment 2 – Object Iteration : Finding the smallest string in a list. Serial streams are about 1.5× slower than loops, but parallel streams again surpass both serial streams and loops, especially as core count increases.

Experiment 3 – Complex Reduction : Summing transaction amounts per user from a list of <userName,price,timeStamp> tuples represented by an Order object. Parallel reduction is slower on a single core but improves markedly with more cores, ultimately beating serial and manual implementations.

Conclusion : For simple traversals, external loops are faster in single‑core scenarios, but parallel streams exploit multi‑core CPUs to deliver superior performance. For complex reductions, parallel streams also provide the best results on multi‑core systems, while serial streams remain comparable to manual code. The article recommends using external loops for trivial tasks, Stream API for complex logic, and parallel streams whenever multiple cores are available.