Mastering Java Garbage Collectors: When to Choose Serial, Parallel, CMS, or G1

Introduction

As Java has evolved, various garbage collectors have emerged, ranging from serial to parallel execution, from high‑throughput to low‑latency designs, aiming to let developers focus on code rather than memory management.

Serial Collector

The older single‑threaded collector pauses application threads during collection. It is simple and efficient. Serial works on the young generation using a mark‑copy algorithm, while Serial Old handles the old generation with a mark‑compact algorithm.

Parallel Collector

Multiple GC threads work in parallel, offering higher throughput on multi‑core CPUs, but application threads still wait. It includes ParNew (young) and Parallel Old (old) collectors, essentially multithreaded versions of Serial.

ParallelScavenge

ParallelScavenge is a young‑generation collector similar to ParNew, using a mark‑copy algorithm. It targets a controllable throughput, with key tuning parameters such as -XX:MaxGCPauseMillis and -XX:GCTimeRatio. Adaptive size policy ( -XX:+UseAdaptiveSizePolicy) can automatically adjust generation sizes.

CMS Collector (Concurrent Mark Sweep)

CMS aims for low pause times and works concurrently with application threads. It performs concurrent marking and sweeping, reducing stop‑the‑world phases. Its four phases are Initial Mark (STW), Concurrent Marking, Remark (STW), and Concurrent Sweep. Advantages include reduced pause times; disadvantages involve higher CPU usage, memory sensitivity, and potential fragmentation.

G1 Collector (Garbage‑First)

G1 introduces a region‑based heap and a pause‑time model, targeting a pause goal (e.g., -XX:MaxGCPauseMillis). It tracks region “value” to prioritize reclamation. G1 includes Young GC, Mixed GC, and Full GC phases, each with specific triggers and steps such as root scanning, RSet updates, object copying, and reference processing.

Key concepts: Regions, Humongous objects, TAMS pointers, Remembered Sets (RSet), and Dirty Card Queues.

Best Practices

Let G1 manage young‑generation size automatically.

Set realistic pause‑time goals; overly aggressive targets increase GC frequency.

Choose CMS for small‑memory workloads and G1 for larger heaps (typically 6‑8 GB).

Conclusion

No single collector is a silver bullet; selection should consider application characteristics and system constraints.