ZGC: A Decade of Innovation by Stefan Johansson

ZGC has been in development for more or less a decade now. This talk will explore the current state of ZGC in JDK 25 and look at how the performance have improved over the years.

1. A ScyllaDB Community
ZGC: JDK 25
Performance Update
Stefan Johansson
OpenJDK GC Engineer

2. A quick look at ZGC

3. The goals of ZGC
Low latency
Pause times below 1 ms
Scalability
Handle TB sized heaps
Auto-tuning
Minimal configuration required

4. ZGC Milestones
JDK 11
Experimental feature

JDK 14
Available on Windows and
macOS
JDK 16
Concurrent stack
scanning
JDK 21
Generational mode (off by default)
enabled with –XX:+ZGenerational
JDK 12
Concurrent class
unloading
JDK 15
Production ready
JDK 17
Automatic scaling of
worker threads
JDK 25
ZGC always generational

5. Current status
Low latency
Pause times below 1 ms
Scalability
Handle TB sized heaps
Auto-tuning
Minimal configuration required

6. Current status
Low latency
•Pause only for synchronization
•Heavy work done concurrently
Scalability
Handle TB sized heaps
Auto-tuning
Minimal configuration required

7. Current status
Low latency
•Pause only for synchronization
•Heavy work done concurrently
Scalability
•Support 16 TB heaps
•Pauses still short
Auto-tuning
Minimal configuration required

8. Current status
Low latency
•Pause only for synchronization
•Heavy work done concurrently
Scalability
•Support 16 TB heaps
•Pauses still short
Auto-tuning
•Little configuration required
•Just set the heap size

9. Performance overview

10. What is good performance?
■It all depends
●Throughput
●Latency
●Footprint

■Trade-offs between those aspects

11. How do ZGC achieve good performance?
■Great low latency alternative
●Basically no pauses

■By doing work concurrently
●Determining liveness
●Moving live objects
●Free memory

■Keep the concurrency overhead low

12. Low concurrency overhead
■Return memory as soon as possible
●Avoid allocation stalls

■Keep CPU overhead at minimum barriers
●Efficient barriers
●Smart algorithms

■Colored pointers

13. Performance details

14. Pause times

15. Pause times

16. Pause times

17. Pause times

18. Pause times

19. Pause times

20. Application throughput

21. Application latency

22. Thank you! Let’s connect.
Stefan Johansson
@kstefanj
kstefanj.github.io