# 打印堆栈信息到标准输出 jstack PID  
# 打印堆栈信息到标准输出，会打印关于锁的信息 jstack -l PID  
强制打印堆栈信息到标准输出，如果使用 jstack PID 没有响应的情况下(此时 JVM 进程可能挂起)，
加 -F 参数 jstack -F PID 

# 下面的命令同等于 jstack PID 
jcmd PID Thread.print  

# 同等于 jstack -l PID 
jcmd PID Thread.print -l 

kill -3 PID 
# 或 kill -QUIT PID 

# JVM 发生 OOM 时，会自动在 /var/log/abc 目录下产生堆 dump 文件 java_pidPID.hprof 
java -XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/var/log/abc/ 

# 将 JVM 的堆 dump 到指定文件，如果堆中对象较多，需要的时间会较长，子参数 format 只支持 b，
即二进制格式
jmap -dump:format=b,file=FILE_WITH_PATH

# 如果 JVM 进程未响应命令，可以加上参数 -F 尝试
jmap -F -dump:format=b,file=FILE_WITH_PATH

# 可以只 dump 堆中的存活对象，加上 live 子参数，但使用 -F 时不支持 live
jmap -dump:live,format=b,file=FILE_WITH_PATH

# -heap 参数用于查看指定 JVM 进程的堆的信息，包括堆的各个参数的值，堆中新生代、年老代的内存大小、使用率等 
jmap -heap PID  

# 同样，如果 JVM 进程未响应命令，可以加上参数 -F 尝试 
jmap -F -heap PID 

Attaching to process ID 68322, please wait
Debugger attached successfully.
Server compiler detected.
JVM version is 25.112-b16

using thread-local object allocation.
Parallel GC with 4 thread(s)

Heap Configuration:
   MinHeapFreeRatio         = 0
   MaxHeapFreeRatio         = 100
   MaxHeapSize              = 268435456 (256.0MB)
   NewSize                  = 8388608 (8.0MB)
   MaxNewSize               = 89128960 (85.0MB)
   OldSize                  = 16777216 (16.0MB)
   NewRatio                 = 2
   SurvivorRatio            = 8
   MetaspaceSize            = 21807104 (20.796875MB)
   CompressedClassSpaceSize = 1073741824 (1024.0MB)
   MaxMetaspaceSize         = 17592186044415 MB
   G1HeapRegionSize         = 0 (0.0MB)

Heap Usage:
PS Young Generation
Eden Space:
   capacity = 41943040 (40.0MB)
   used     = 1701504 (1.6226806640625MB)
   free     = 40241536 (38.3773193359375MB)
   4.05670166015625% used
From Space:
   capacity = 4194304 (4.0MB)
   used     = 0 (0.0MB)
   free     = 4194304 (4.0MB)
   0.0% used
To Space:
   capacity = 5242880 (5.0MB)
   used     = 0 (0.0MB)
   free     = 5242880 (5.0MB)
   0.0% used
PS Old Generation
   capacity = 30408704 (29.0MB)
   used     = 12129856 (11.56793212890625MB)
   free     = 18278848 (17.43206787109375MB)
   39.889421134159484% used

16658 interned Strings occupying 1428472 bytes.

获取堆中的类实例统计

# 打印 JVM 堆中的类实例统计信息，以占用内存的大小排序，同样，如果 JVM 未响应命令，也可以使用 -F 参数 
jmap -histo PID  

# 也可以只统计堆中的存活对象，加上 live 子参数，但使用 -F 时不支持 live 
jmap -histo:live PID 

# 等同 jmap -dump:live,format=b,file=FILE_WITH_PATH
jcmd PID GC.heap_dump FILE_WITH_PATH

# 等同 jmap -dump:format=b,file=FILE_WITH_PATH
jcmd PID GC.heap_dump -all FILE_WITH_PATH

# 等同 jmap -histo:live PID
jcmd PID GC.class_histogram

# 等同 jmap -histo PID
jcmd PID GC.class_histogram -all

In this video I explain some 21 JVM parameters which are suited for most server applications. If you have any questions, you can read those links below for more information or just ask in the comments section.

I run several Java enterprise server applications. I often wondered – what are the best „default“ JVM settings for a server application to start with in production? I read a lot on the web and tried several things myself and wanted to share what I found out, so far. Links containing more information about JVM optimization can be found here:

http://blog.sokolenko.me/2014/11/javavm-options-production.html

http://www.petefreitag.com/articles/gctuning/

http://stas-blogspot.blogspot.de/2011/07/most-complete-list-of-xx-options-for.html

So let’s start:

-server

Use „-server“: All 64-bit JVMs use the server VM as default anyway. This setting generally optimizes the JVM for long running server applications instead of startup time. The JVM will collect more data about the Java byte code during program execution and generate the most efficient machine code via JIT.

-Xms=<heap size>[g|m|k] -Xmx=<heap size>[g|m|k]

The „-Xmx/-Xms“ settings specify the maximum and minimum values for the JVM heap memory. For servers, both params should have the same value to avoid heap resizing during runtime. I’ve applications running with 16GB heap sizes without an issue.

Depending on your application, you will have to try out how much memory will be best suited for your use case.

-XX:MaxMetaspaceSize=<metaspace size>[g|m|k]

Java 8 has no „Permanent Generation“ (PermGen) anymore but requires additional „Metaspace“ memory instead. This memory is used, in addition to the heap memory we specified before, for storing class meta data information.

The default size will be unlimited – I tend to limit MaxMetaspaceSize with a somewhat high value. Just in case something goes wrong with the application, the JVM will not hog all the memory of the server.

I suggest: Let your application run for a couple of days to get a feeling for how much Metaspace Size it uses normally. Upon next restart of the application set the limit to e.g. double the value.

-XX:+CMSClassUnloadingEnabled

Additionally, you might want to allow the JVM to unload classes which are held in memory but no code is pointing to them any more. If your application generates lots of dynamic classes, this is what you want.

-XX:+UseConcMarkSweepGC

This option makes the JVM use the ConcurrentMarkSweepGC – It can do much work in parallel to program execution but in some circumstances a „full GC“ with a „STW pause“ might still occur. I’ve read many articles and came to the conclusion that this GC is still the best one for server workloads.

-XX:+CMSParallelRemarkEnabled

The option CMSParallelRemarkEnabled means the remarking is done in parallel to program execution – which is what you want if your server has many cores (and most servers do).

 -XX:+UseCMSInitiatingOccupancyOnly  -XX:CMSInitiatingOccupancyFraction=<percent>

Normally the GC will use heuristics to know when it’s time to clear memory. GC might kick in too late with default settings (causing full-Gcs).
Some sources say it might be a good idea to disable heuristics altogether and just use generation occupancy to start a CMS collection cycle. Setting values around 70% worked fine for all of my applications and use cases.

-XX:+ScavengeBeforeFullGC

The first option tells the GC to first free memory by clearing out the „young generation“ or fairly new objects before doing a full GC.

-XX:+CMSScavengeBeforeRemark

CMSScavengeBeforeRemark does attempt a minor collection before the CMS remark phase – thus keeping the remark pause afterwards short.

-XX:+CMSClassUnloadingEnabled

The option „-XX:+CMSClassUnloadingEnabled“ here tells the JVM to unload classes, which are not needed any more by the running application. If you deploy war files to an application server like wildfly, tomcat or glassfish without restarting the server after the deployment, this flag is for you.

-XX:+ExplicitGCInvokesConcurrentAndUnloadsClasses

The option „-XX:+ExplicitGCInvokesConcurrentAndUnloadsClasses“ is especially important if your application uses RMI (remote method invocation). The usage of RMI will cause the JVM to do a FULL-GC EVERY HOUR! This might be a very bad idea for large heap sizes because the FULL-GC pause might take up to several seconds. It would be better to do a concurrent GC and try to unload unused classes to free up more memory – which is exactly what the second option does.

-XX:+PrintGCDateStamps -verbose:gc -XX:+PrintGCDetails -Xloggc:"<path to log>"

These options shown here will write out all GC related information to a specified log file. You can see how well your GC configuration works by looking into it.

I personally prefer to use the „Visual GC“ plug in for the „Visual VM“ tool to monitor the general JVM and GC behavior.

-XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=<path to dump>`date`.hprof

When your JVM runs out of memory, you will want to know why. Since the OOM error might be hard to reproduce and you want to get your production server up and running again – you should specify a path for a heap dump. When things have settled down, you can analyze the dump afterwards.

-Djava.rmi.server.hostname=<external IP> -Dcom.sun.management.jmxremote.port=<port>

These options will help you to specify an IP and port for JMX – you will need those ports open to connect remotely to a JVM running on a server for tools like VisualVM. You can gain deep insights over cpu and memory usage, gc behaviour, class loading, thread count and usage of your application this way.

Lastly, I would like to recommend to you the VisualVM tool which is bundled with the Java 8 JDK. You can use it to gain more insights about your specific application behaviour on the JVM – like cpu and memory usage, thread utilisation and much more.

VisualVM can be extended with a plug in called „Visual GC“. It will briefly show you VERY detailed information about the usage of the young and old generation object spaces. You can easily spot problems with garbage collection simply by analyzing these graphs during application runtime.

Thank you very much for watching! If you liked the video you might consider giving it a „thumbs up“. If you have any questions – just put them in the comments section. I will reply as quickly as possible.

-------------------------------------------------------

-XX:+UseCompressedOops [If Max Heap allocation is less than 32GB]
This can save a significant amount of memory and this option should already be enabled by default on recent java 8 versions. This option allowes object references to be stored as 32-bit values instead of 64-bit on 64-bit JVMs. This leads to before mentioned memory savings.

-XX:+AggressiveOpts
This option will enable performance options which are hoped to become enabled by default in upcoming released of the JVM. This option sets some performance settings but is marked as experimental! So you should only enable it, when you have to possibility to test your application thoroughly before enabling this flag on an production server.

-XX:+UseStringDeduplication
Since Java 8 update 20 you can use this option to reduce the memory usage of your application. The JVM will spot identical strings in memory, remove the duplicated and point all references to the remaining, single instance of the string.

-XX:+UseG1GC
Will tell the JVM to use the most recent G1 garbage collector. You are trading better application response times (due to shorter gc times with G1) against lower throughput (compared against good old ConcMarkSweepGC / CMS). If your application can deliver more value through short gc times, then G1 is definately better suited. Otherwise on Java 8, I’d recommend sticking with CMS.

Concerning your Tomcat 8 question, I’d suggest you have a look into it with the „VisualVM“ tool. Look at memory usage, GC times (visual GC plugin), pull and analyse stack traces or thread dumps to find the weak spot. You might also consider attaching a debugger to tomcat to find the bug.

https://www.maknesium.de/21-most-important-java-8-vm-options-for-servers