zgc-jdk25-benchmarks/README.md

Generational ZGC vs G1 on JDK 25 — runnable benchmarks and internals

Companion repository for Generational ZGC on JDK 25: Benchmarks vs G1 on ankurm.com.

The blog post covers the concepts most people need. This repository covers everything — every flag, every failure mode, every measurement pitfall, with runnable code, real captured output, and the reasoning behind each choice.

Everything here was compiled and executed on java 25.0.3+9-LTS-195. Raw, unedited output for every run is in results/.

---

Quick start

git clone https://ankurm.com/git.app/asmhatre/zgc-jdk25-benchmarks.git
cd zgc-jdk25-benchmarks

# Windows
pwsh scripts/run-all.ps1 -JavaHome 'C:\Program Files\Java\jdk-25.0.3'

# Linux / macOS
./scripts/run-all.sh /path/to/jdk-25

Or run a single piece — nothing but a JDK 25 is required for anything outside jmh/:

javac -d out $(find latency tuning internals analysis -name '*.java')

java -XX:+UseZGC  -Xms2g -Xmx2g -cp out com.ankurm.zgc.latency.LatencyHarness
java -XX:+UseG1GC -Xms2g -Xmx2g -cp out com.ankurm.zgc.latency.LatencyHarness

---

Headline result

60 seconds, 20,000 req/s open loop, 2 GB heap, ~585 MB live set, ~328 MB/s allocation rate. Full output: results/03-latency-zgc.txt, results/03-latency-g1.txt.

Metric	Generational ZGC	G1
p50 response time	0.006 ms	0.005 ms
p90	0.009 ms	0.009 ms
p99	0.029 ms	8.042 ms
p99.9	1.437 ms	95.169 ms
p99.99	23.675 ms	126.949 ms
max	32.178 ms	133.018 ms
total stop-the-world time	1.503 ms	1,139.624 ms
longest single pause	0.054 ms	104.449 ms
Full GCs / evacuation failures	0 / 0	4 / 16

G1 wins the median. ZGC wins everything past p99. The two are indistinguishable up to p90 — which is the part most "ZGC is faster" posts leave out.

---

What is in here

Code

Module	What it demonstrates	Deps
env/Env.java	Environment capture; proves generational mode from JMX bean names	none
latency/	Open-loop latency harness with coordinated-omission accounting	none
tuning/AllocationStallDemo.java	Provokes and measures ZGC allocation stalls via JFR	none
internals/PageSizeDemo.java	ZGC page classes vs G1 humongous, read from the live VM	none
analysis/GcLogParser.java	Unified-log parser that keeps pauses and concurrent phases apart	none
jmh/	Throughput and barrier microbenchmarks	JMH 1.37

JMH benchmarks

Class	Isolates
AllocationBench	TLAB allocation path at 64 B / 1 KB / 32 KB
LoadBarrierBench	ZGC's load barrier — pointer chase vs primitive sum control
StoreBarrierBench	Write barriers — old→young vs null vs old→old vs primitive control
PromotionBench	Medium-lived objects; the case where "most objects die young" fails

Documentation

Chapter	Covers
1. The generational model	JEP 439 → 474 → 490 timeline, what ZGenerational does now, why there is no -Xmn
2. Barriers	Colored pointers, load barrier, store barrier, remembered sets, ZBufferStoreBarriers
3. Allocation stalls	The ZGC failure mode pause charts cannot see, and the 10 ms JFR threshold that hides it
4. Tuning reference	Every ZGC product + diagnostic flag, and the G1 flags that still matter
5. Logging and JFR	Reading -Xlog:gc*, the JMX bean trap, the JFR events worth enabling
6. Methodology	Coordinated omission, timer granularity, the JMH -jvmArgs fork trap
7. Corner cases	Compressed oops, large-page fragmentation, virtual thread stacks, containers, System.gc()
8. Choosing a collector	A decision procedure, sizing guidance, when the answer is neither

---

Six findings you will not read elsewhere

1. G1's coordinated-omission gap on this workload is 2910x. A closed-loop harness reports G1's p99.9 as 0.033 ms. It is really 95.169 ms. → docs/06-methodology.md §6.1
2. ZGC disables compressed oops. Every reference is 8 bytes instead of 4, so a reference-dense heap can grow 30–40% on switching collector, before any latency benefit. → docs/07-corner-cases.md §7.1
3. jdk.ZAllocationStall has a 10 ms default JFR threshold. In a run whose mean stall was 4.33 ms, a default profile discards most of them. → docs/03-allocation-stalls.md §3.3
4. new byte[4*1024*1024] gets a ZGC Large page, not a Medium one. The 16-byte array header pushes it over the medium-object limit, and it consumes an 8 MB page. → docs/07-corner-cases.md §7.2
5. SoftMaxHeapSize and ZAllocationSpikeTolerance did nothing under sustained overload. Measured: 32,200 / 31,967 / 32,442 stalls across baseline and both mitigations. They buy headroom for spikes; they cannot conjure heap that does not exist. → results/04-stalls-zgc-softmax.txt
6. The throughput ranking inverts under sustained promotion. G1 is 6.6% faster with no survivors; ZGC is 10.4% faster with 262,144 of them. "G1 wins throughput" only holds on the workload that suits generational collection best. → results/07-jmh-zgc.txt

---

Reference machine

CPU	AMD Ryzen 5 5600U, 6 cores / 12 threads
RAM	15.3 GB
OS	Windows 11 Home
JDK	java 25.0.3+9-LTS-195 (Oracle)
Heap	-Xms2g -Xmx2g for all latency runs; -Xms512m -Xmx512m for stall runs

A laptop, not a server. The absolute numbers are illustrative; the relationships are what generalise. Run it on your own hardware — that is what it is for.

---

Licence

MIT. See LICENSE.