JVM Garbage Collectors Compared — G1, ZGC, Shenandoah, Parallel | Jvm Internals Interview Q&A | InterviewExplainer

Q: JVM Garbage Collectors Compared — G1, ZGC, Shenandoah, Parallel

**G1GC** (default Java 9+) targets **~200ms pauses** with region-based collection. **ZGC** and **Shenandoah** trade **CPU** for **sub-millisecond** pauses on large heaps. **Parallel GC** maximizes **throughput** with longer **stop-the-world** pauses — best for **batch** jobs. Pick by **latency SLA**, not brand.

JVM Garbage Collectors Compared — G1, ZGC, Shenandoah, Parallel

Fast recall

Quick revision

Core answer

G1GC (default Java 9+) targets ~200ms pauses with region-based collection. ZGC and Shenandoah trade CPU for sub-millisecond pauses on large heaps. Parallel GC maximizes throughput with longer stop-the-world pauses — best for batch jobs. Pick by latency SLA, not brand.

Java ships with four main garbage collectors, each making a different trade-off between throughput, pause duration, and CPU overhead. There is no universally best choice — the right GC algorithm depends on what your application can tolerate: long pauses for high throughput, or short pauses at higher CPU cost.

Parallel GC:

Highest throughput — throws all threads at collection during stop-the-world pauses
Pauses can last seconds — suitable for batch processing and offline analytics, not for user-facing services

G1GC (default since Java 9):

Region-based heap — collects regions with the highest garbage ratio first
Soft pause-time target via -XX:MaxGCPauseMillis=200; adjusts collection scope to meet it
Avoids fragmentation issues that killed CMS; good default for most web services

ZGC:

Almost all work done concurrently with application threads — sub-millisecond pauses regardless of heap size
Higher CPU usage for concurrent GC work
Use when even 200ms pauses are unacceptable: algorithmic trading, real-time pricing, strict p99 SLAs

Common mistake:

Assuming ZGC is universally better — on small heaps under 4GB, G1 pauses are already short; ZGC's CPU overhead provides no net benefit

I recommend G1 as the default, enabling GC logging, and only switching if p99 latency analysis shows GC pauses as the dominant contributor.

Interview-ready guide

Interview answer

Parallel GC:

Highest throughput — throws all threads at collection during stop-the-world pauses
Pauses can last seconds — suitable for batch processing and offline analytics, not for user-facing services

G1GC (default since Java 9):

Region-based heap — collects regions with the highest garbage ratio first
Soft pause-time target via -XX:MaxGCPauseMillis=200; adjusts collection scope to meet it
Avoids fragmentation issues that killed CMS; good default for most web services

ZGC:

Almost all work done concurrently with application threads — sub-millisecond pauses regardless of heap size
Higher CPU usage for concurrent GC work
Use when even 200ms pauses are unacceptable: algorithmic trading, real-time pricing, strict p99 SLAs

Common mistake:

Assuming ZGC is universally better — on small heaps under 4GB, G1 pauses are already short; ZGC's CPU overhead provides no net benefit

I recommend G1 as the default, enabling GC logging, and only switching if p99 latency analysis shows GC pauses as the dominant contributor.

03 · Deep dive