Image Modified

Contents

Supported Platforms

Platform	Supported	Since	Comment
Linux/x64		JDK 15 (Experimental since JDK 11)
Linux/AArch64		JDK 15 (Experimental since JDK 13)
Linux/PowerPC		JDK 18
macOS/x64		JDK 15 (Experimental since JDK 14)
macOS/AArch64		JDK 17
Windows/x64		JDK 15 (Experimental since JDK 14)	Requires Windows version 1803 (Windows 10 or Windows Server 2019) or later.
Windows/AArch64		JDK 16

Supported Platforms

ZGC is currently available on:

• Linux/x86_64
• Solaris/SPARC (M7/T7 and later)
  

Support for other platforms might be added in the future, if there is enough demand.

Download and Build

The source code can be found in the ZGC project repository. Just clone, configure and make.

$ hg clone http://hg.openjdk.java.net/zgc/zgc
$ cd zgc
$ sh configure
$ make images

This will build a complete JDK for you, which includes ZGC. On Linux, the root directory of the new JDK will be found here:

./build/linux-x86_64-normal-server-release/images/jdk

And the Java launcher will be found in its usual place, here:

./build/linux-x86_64-normal-server-release/images/jdk/bin/java

Quick Start

If you're trying out ZGC for the first time, start by using the following GC options:

-XX:+UseZGC -Xms<size> -Xmx<size> -Xlog:gc

For more logging and basic tuning to improve throughput and latencydetailed logging, use the following options:

-XX:+UseZGC -Xms<size> -Xmx<size> -Xlog:gc* 

Note for JDK 21 you need to explicitly use the following options to use the newer, generational mode of ZGC:

-XX:+UseLargePagesUseZGC -XX:ParallelGCThreads=<n> -XX:ConcGCThreads=<n>+ZGenerational 

See below for more information on these and additional options.

Configuration & Tuning

Enabling ZGC

Use the -XX:+UseZGC option to enable ZGC.

Setting Heap Size

In general, ZGC works best when there is enough heap headroom to keep up with the allocation rate without having to run back to back GCs. With too little headroom the GC will simply not be able to keep up and Java threads will be stalled to allow the GC to catch up. Stalling of Java threads should basically be avoided at all cost by increasing the heap size (or as a secondary option, increase the number of concurrent GC threads, see below).

For optimal performance, setting the initial heap size (-Xms) equal to the maximum heap size (-Xmx) is generally recommended.

Setting Parallel/Concurrent GC Threads

ZGC uses both -XX:ParallelGCThreads=<threads> and -XX:ConcGCThreads=<threads> to determine how many worker threads to use during different GC phases. If they are not set, ZGC will try to select an appropriate number. However, please note that the optimal number of threads to use heavily depends on the characteristics of the workload you're running, which means that you almost always want to explicitly specify these to get optimal throughput and latency. We hope to be able to remove this recommendation at some point in the future, when ZGC's heuristics for this becomes good enough, but for now it's recommended that you try different settings and pick the best one.

ParallelGCThreads sets the level of parallelism used during pauses and hence directly affects the pause times. Generally speaking, the more threads the better, as long as you don't over provision the machine (i.e. use more threads than cores/hw-threads) or the application root set is so small that is can easily be handled by just a few threads.

The following GC phases are affected by ParallelGCThreads:

• Pause Mark Start - Number of threads used for marking roots.
• Pause Mark End - Number of threads used for weak root processing (StringTable, JNI Weak Handles, etc.).
• Pause Relocate Start - Number of threads used for relocating roots.

ConcGCThreads sets the level of parallelism used during concurrent phases. The number of threads to use during these phases is a balance between allowing the GC to make progress and not stealing too much CPU time from the application. Generally speaking, if there are unused CPUs/cores in the system, always allow concurrent threads to use them. If the application is already using all CPUs/cores, then the machine is essentially already over-provisioned and you have to allow for a throughput reduction by either letting concurrent GC threads steal/compete for CPU time, or by actively reducing the application CPU footprint.

ZGC has been designed to be adaptive and to require minimal manual configuration. During the execution of the Java program, ZGC dynamically adapts to the workload by resizing generations, scaling the number of GC threads, and adjusting tenuring thresholds. The main tuning knob is to increase the maximum heap size.

Overview

The following JVM options can be used with ZGC:

In addition to these the following flags are available when the generational mode is enabled with -XX:+UseZGC -XX:+ZGenerational:

Setting Heap Size

The most important tuning option for ZGC is setting the maximum heap size, which you can set with the -Xmx command-line option. Because ZGC is a concurrent collector, you must select a maximum heap size such that the heap can accommodate the live-set of your application and there is enough headroom in the heap to allow allocations to be serviced while the GC is running. How much headroom is needed very much depends on the allocation rate and the live-set size of the application. In general, the more memory you give to ZGC the better. But at the same time, wasting memory is undesirable, so it’s all about finding a balance between memory usage and how often the GC needs to run.

ZGC has another command-line option related to the heap size named -XX:SoftMaxHeapSize. It can be used to set a soft limit on how large the Java heap can grow. ZGC will strive to not grow beyond this limit, but is still allowed to grow beyond this limit up to the maximum heap size. ZGC will only use more than the soft limit if that is needed to prevent the Java application from stalling and waiting for the GC to reclaim memory. For example, with the command-line options -Xmx5g -XX:SoftMaxHeapSize=4g ZGC will use 4GB as the limit for its heuristics, but if it can't keep the heap size below 4GB it is still allowed to temporarily use up to 5GB.

Setting Concurrent GC Threads

Note! This section pertain to the non-generational version of ZGC. Generational ZGC has a more adaptive implementation and you are less likely to need to tweak the GC threads.

The second tuning option one might want to look at is setting the number of concurrent GC threads (-XX:ConcGCThreads=<number>). ZGC has heuristics to automatically select this number. This heuristic usually works well but depending on the characteristics of the application this might need to be adjusted. This option essentially dictates how much CPU-time the GC should be given. Give it too much and the GC will steal too much CPU-time from the application. Give it too little, and the application might allocate garbage faster than the GC can collect it.

NOTE!! Starting from JDK 17, ZGC dynamically scales up and down the number of concurrent GC threads. This makes it even more unlikely that you'd need to adjust the concurrent number of GC threads.

NOTE!!! NOTE! In general, if low latency (i.e. low application response time) is important to for you application, then never over-provision your system. Ideally, your system should never have more than 70% CPU utilization.

The following GC phases are affected by ConcGCThreads:

• Concurrent Mark - Number of threads used for concurrent marking.
• Concurrent Reference Processing - Number of threads used for concurrent reference processing (i.e. handling Soft/Weak/Final/PhantomReference objects).
• Concurrent Relocate - Number of threads used for concurrent relocation.

Example:

When running SPECjbb2015, on a two socket Intel Xeon E5-2690 machine, which a total of 2 x 8 = 16 cores (with hyper-threading, 2 x 16 = 32 HW-threads) using a 128G heap, the following options typically results in optimal throughput and latency:

-XX:+UseZGC -Xms128G -Xmx128G -XX:+UseLargePages -XX:ParallelGCThreads=20 -XX:ConcGCThreads=4

Returning Unused Memory to the Operating System

By default, ZGC uncommits unused memory, returning it to the operating system. This is useful for applications and environments where memory footprint is a concern, but might have a negative impact on the latency of Java threads. You can disable this feature with the command-line option -XX:-ZUncommit. Furthermore, memory will not be uncommitted so that the heap size shrinks below the minimum heap size (-Xms). This means this feature will be implicitly disabled if the minimum heap size (-Xms) is configured to be equal to the maximum heap size (-Xmx).

You can configure an uncommit delay using -XX:ZUncommitDelay=<seconds> (default is 300 seconds). This delay specifies for how long memory should have been unused before it's eligible for uncommit.

NOTE! Allowing the GC to commit and uncommit memory while the application is running could have a negative impact on the latency of Java threads. If extremely low latency is the main reason for running with ZGC, consider running with the same value for -Xmx and -Xms, and use -XX:+AlwaysPreTouch to page in memory before the application starts.

NOTE!! On Linux, uncommitting unused memory requires fallocate(2) with FALLOC_FL_PUNCH_HOLE support, which first appeared in kernel version 3.5 (for tmpfs) and 4.3 (for hugetlbfs).

Using

Large Pages

Configuring ZGC to use large pages will generally yield better performance (in terms of throughput, latency and start up time) and comes with no real disadvantage, except that it's slightly more complicated to setup. The setup process typically requires root privileges, which is why it's not enabled by default.

Enabling Large Pages On Linux

On Linux x86, large pages ( pages are also known as "huge pages" on Linux/x86 and ) have a size of 2MB.

Let's assume you want a 16G 16GB Java heap. That means you need 16G 16GB / 2M 2MB = 8192 huge pages.

First assign The heap requires at least 16G 16GB (8192 pages) of memory to the pool of huge pages. The "at least" part is important, since enabling the use of large pages in the JVM means that not only the GC will try to use these for the Java heap, but also that heap along with other parts of the JVM will try to use them large pages for various internal data structures (such as code heap , and marking bitmaps, etc). In this example we you will therefore reserve 9216 pages (18G18GB) to allow for 2G 2GB of non-Java heap allocations to use large pages.

Configure the system's huge page pool to have the required number of pages (requires root privileges):

$ echo 9216 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages

Note that the above command is not guaranteed to be successful if the kernel can not cannot find enough free huge pages to satisfy the request. Also note that it might take some time for the kernel to process the request. Before proceeding, check the number of huge pages assigned to the pool to make sure the request was successful and has completed.

$ cat /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
9216 9216

NOTE! If you're using a Linux kernel >= 4.14, then the next step (where you mount a hugetlbfs filesystem) can be skipped. However, if you're using an older kernel then ZGC needs to access large pages through a hugetlbfs filesystem.

Mount a hugetlbfs filesystem (requires root privileges) and make it accessible to the user running the JVM (in this example we're assuming this user has 123 as its uid).

$ mkdir /hugepages
$ mount -t hugetlbfs -o uid=123 nodev /hugepages 

Now start the JVM using the -XX:+UseLargePages option.

$ java -XX:+UseZGC -Xms16G -Xmx16G -XX:+UseLargePages ...

If there are more than one accessible hugetlbfs filesystem available, then (and only then) do you also have to use -XX:ZPathAllocateHeapAt to specify the path to the filesystems you want to use. For example, assume there are multiple accessible hugetlbfs filesystems mounted, but the filesystem you specifically want to use it mounted on /hugepages, then use the following options.

$ java -XX:+UseZGC -Xms16G -Xmx16G -XX:+UseLargePages -XX:ZPathAllocateHeapAt=/hugepages ...

NOTE! The configuration of the huge page pool and the mounting of the hugetlbfs file system is not persistent across reboots, unless adequate measures are taken.

Enabling Transparent Huge Pages

On Linux

NOTE! On Linux, using ZGC with transparent huge pages enabled requires kernel >= 4.7.'

Use the following options to enable transparent huge pages in the VM:

-XX:+UseLargePages -XX:+UseTransparentHugePages

These options tell the JVM to issue madvise(..., MADV_HUGEPAGE) calls for memory it maps, which is useful when using transparent huge pages in madvise mode.

To enable transparent huge pages, you also need to have the appropriate mode set in configure the kernel by enabling madvise mode.

$ echo madvise > /sys/kernel/mm/transparent_hugepage/enabled

ZGC uses shmem huge pages for the heap, so the following kernel setting also needs to be configured:

$ echo advise > /sys/kernel/mm/transparent_hugepage/shmem_enabled

It is important to check these kernel settings when comparing the performance of different GCs. Some Linux distributions forcefully enable transparent huge pages for private pages by configuring /sys/kernel/mm/transparent_hugepage/enabled to be set to always, while leaving /sys/kernel/mm/transparent_hugepage/shmem_enabled at the default never. In this case all GCs but ZGC will make use of transparent huge pages for the heap. See Transparent Hugepage Support for more information. See the kernel documentation for more information.The options above tells the JVM to issue madvise(..., MADV_HUGEPAGE) calls for memory it mapps, which is useful when using transparent huge pages in madvise mode.

Enabling NUMA Support

ZGC has basic NUMA support, which means it will try it's best to direct Java heap allocations to NUMA-local memory. This feature is enabled by default. However, it will automatically be disabled if the JVM detects that it's bound to a sub-set of the CPUs in the systemonly use memory on a single NUMA node. In general, you don't need to worry about this setting, but if you want to explicitly override the JVM's decision you can do so by using the -XX:+UseNUMA or -XX:-UseNUMA options.

When running on a NUMA machine (e.g. a multi-socket x86 machine), having NUMA support enabled will often give a noticeable performance boost.

Enabling GC Logging

GC logging is enabled using the following command-line option:

-Xlog:<tag set>,[<tag set>, ...]:<log file>

For general information/help on this option:

-Xlog:help

To enable basic logging (one line of output per GC):

-Xlog:gc:gc.log

To enable GC logging that is useful for tuning/performance analysis:

-Xlog:gc*:gc.log

Where gc* means log all tag combinations that contain the gc tag, and :gc.log means write the log to a file named gc.log.

Change Log

JDK 24

• Removed the non-generational mode of ZGC (JEP 490)

JDK 23

• Make Generational ZGC the default ZGC version (and deprecate non-generational ZGC) (JEP 474)
• Latency Issue Due to ICBufferFull Safepoints Resolved (JDK-8322630) 

JDK 21

• Support for generations (-XX:+ZGenerational) (JEP 439)

JDK 18

• Support for String Deduplication (-XX:+UseStringDeduplication)
• Linux/PowerPC support
• Various bug-fixes and optimizations

JDK 17

• Dynamic Number of GC threads
• Reduced mark stack memory usage
• macOS/aarch64 support
• GarbageCollectorMXBeans for both pauses and cycles
• Fast JVM termination

JDK 16

• Concurrent Thread Stack Scanning (JEP 376)
• Support for in-place relocation
• Performance improvements (allocation/initialization of forwarding tables, etc)

JDK 15

• Production ready (JEP 377)
• Improved NUMA awareness
• Improved allocation concurrency
• Support for Class Data Sharing (CDS)
• Support for placing the heap on NVRAM
• Support for compressed class pointers
• Support for incremental uncommit
• Fixed support for transparent huge pages
• Additional JFR events

JDK 14

• macOS support (JEP 364)
• Windows support (JEP 365)
• Support for tiny/small heaps (down to 8M)
• Support for JFR leak profiler
• Support for limited and discontiguous address space
• Parallel pre-touch (when using -XX:+AlwaysPreTouch)
• Performance improvements (clone intrinsic, etc)
• Stability improvements

JDK 13

• Increased max heap size from 4TB to 16TB
• Support for uncommitting unused memory (JEP 351)
• Support for -XX:SoftMaxHeapSIze
• Support for the Linux/AArch64 platform
• Reduced Time-To-Safepoint

JDK 12

• Support for concurrent class unloading
• Further pause time reductions

JDK 11

• Initial version of ZGC
• Does not support class unloading (using -XX:+ClassUnloading has no effect)

FAQ

What does the "Z" in ZGC stand for?

It doesn't stand for anything, ZGC is just a name. It was originally inspired by, or a homage to, ZFS (the filesystem) which in many ways was revolutionary when it first came out. Originally, ZFS was an acronym for "Zettabyte File System", but that meaning was abandoned and it was later said to not stand for anything. It's just a name. See Jeff Bonwick's Blog for more details.

Is it pronounced "zed gee see" or "zee gee see"?

There's no preferred pronunciation, both are fine.