Java on Linux for Devs and Ops Alexey Ragozin 

Java Vs Linux Java VM Managed memory  Garbage collection Multithreading Cross platform API  File system  Networking Linux – User space VM Memory management  Virtual memory Permissive multitasking File system and networking 

Memory 

Java Memory Java Heap Young Gen Old Gen Perm Gen Non-Heap JVM Memory Thread Stacks NIO Direct Buffers Metaspace Compressed Class Space Code Cache Native JVM Memory Non-JVM Memory (native libraries) Java 7 Java 8 Java 8 -Xms/-Xmx -Xmn -XX:PermSize -XX:MaxDirectMemorySize -XX:ReservedCodeCacheSize -XX:MaxMetaspaceSize -XX:CompressedClassSpaceSize Java Process Memory -XX:ThreadStackSize per thread 

Linux memory Memory is managed in pages (4k) on x86 / AMD64 (Huge page support exist on Linux, but it has own problems) Pages from process point of view - Virtual address reservation - Committed memory page - File mapped memory page 

Linux memory Pages from OS point of view Private Shared Anonymous File backed Shared memory Private process memory Executables / Libraries Memory mapped files Memory mapped files Cache / Buffers https://techtalk.intersec.com/2013/07/memory-part-1-memory-types/ 

Understanding memory metrics 

Understanding memory metrics OS Memory  Memory Used/Free – misleading metric  Swap used – should be zero  Buffers/Cached – essentially this is free memory* Process  VIRT – address space reservation - not a memory!  RES – resident size - key memory footprint  SHR – shared size 

Understanding memory metrics  Buffers – pages used for non-file disk data (e.g. file system metadata)  Cached – pages mapped to file data Non-dirty pages – are essentially free memory. Such pages can be used immediately to fulfill memory allocation request. Dirty pages – writable file mapped pages which has modifications not synchronized to disk. 

Linux Process Memory Summary Resident Commited Virtual Zeroed Pages + Swap 

Java Memory Facts Swapping intolerance  GC does heap wide scans  Any Java thread blocked by page fault can block Stop the World pause Java never give up memory to OS  Yes, G1 and serial collector can give memory back to OS  In practice, JVM would still hold all memory it is allowed too 

Out of Memory in Java public void doWork() { Object[] hugeArray = new Object[HUGE_SIZE]; for(int i = 0; i != hugeArray.length; ++i) { hugeArray[i] = calc(i); } } 

Out of Memory in Linux public void doWork() { Object[] hugeArray = new Object[HUGE_SIZE]; for(int i = 0; i != hugeArray.length; ++i) { hugeArray[i] = calc(i); } } 

JVM Out of Memory JVM heap is full and –Xmx limit reached  Start Full GC  If reclaimed memory below threshold throw OutOfMemoryError  OOM error is not recoverable, useful to shutdown gracefully  -XX:OnOutOfMemoryError="kill -9 %p“  OOM can be caught and discarded prolonging agony 

JVM Out of Memory JVM heap is full and at –Xmx limit JVM heap is full but below –Xmx limit  Heap is extended by requesting more memory from OS  If OS rejects memory requests JVM would crash (no OOM error) 

JVM Out of Memory JVM heap is full and at –Xmx limit JVM heap is full but below –Xmx limit NIO direct buffers capacity is capped by JVM  -XX:MaxDirectMemorySize=16g  Cap is enfored by JVM  OOM error in case is limit has been reached – recoverable 

Sizing Java Process Live set: test empirically Young space size: control GC frequency (G1 collector manages young size automatically) Heap size: young space + live set + reserve Reserve: 30% - 50% of live set OS memory footprint > Java Heap Size 

Java in Docker  Guest resources are capped via Linux cgroups  Kernel memory pools can be limited resident / swap / memory mapped  Limits are global for container  Resources restrictions violations remediated by container termination Plan your container size carefully! 

ulimits > ulimit -a core file size (blocks, -c) 1 data seg size (kbytes, -d) unlimited scheduling priority (-e) 0 file size (blocks, -f) unlimited pending signals (-i) 4134823 max locked memory (kbytes, -l) 64 max memory size (kbytes, -m) 449880520 open files (-n) 1024 pipe size (512 bytes, -p) 8 POSIX message queues (bytes, -q) 819200 real-time priority (-r) 0 stack size (kbytes, -s) 8192 cpu time (seconds, -t) unlimited max user processes (-u) 4134823 virtual memory (kbytes, -v) 425094640 file locks (-x) unlimited May prevent you form starting large JVM Core dump disabled Has no effect on Linux 

Threads 

Java Threads Java threads are normal OS threads  Each Java thread are mapped to Linux thread  Java code shares stack with native code  You can use many native Linux tools for diagnostic 

Java Threads in ps ragoale@axcord02:~> ps -T -p 6857 -o pid,tid,%cpu,time,comm PID TID %CPU TIME COMMAND 6857 6857 0.0 00:00:00 java 6857 6858 0.0 00:00:00 java 6857 6859 0.0 00:00:16 java 6857 6860 0.0 00:00:16 java 6857 6861 0.0 00:00:18 java 6857 6862 0.1 00:13:05 java 6857 6863 0.0 00:00:00 java 6857 6864 0.0 00:00:00 java 6857 6877 0.0 00:00:00 java 6857 6878 0.0 00:00:00 java 6857 6880 0.0 00:00:20 java 6857 6881 0.0 00:00:04 java 6857 6886 0.0 00:00:00 java 6857 6887 0.0 00:03:07 java ... This thread mapping is “typical” and not accurate, use jstack to get Java thread information for thread ID VM Operation Thread GC Threads Other application and JVM threads 

Java Thread in jstack jstack (JDK tool) Full thread dump Java HotSpot(TM) 64-Bit Server VM (25.60-b23 mixed mode): "Attach Listener" #65 daemon prio=9 os_prio=0 tid=0x0000000000cbc800 nid=0x1f0 waiting on condition [0x0000000000000000] java.lang.Thread.State: RUNNABLE "pool-1-thread-20" #64 prio=5 os_prio=0 tid=0x00000000009d5000 nid=0x1c04 waiting on condition [0x00007fa109e55000] java.lang.Thread.State: WAITING (parking) at sun.misc.Unsafe.park(Native Method) - parking to wait for <0x00000000d3ab9e50> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject) at java.util.concurrent.locks.LockSupport.park(LockSupport.java:175) at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:2039) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:1088) at java.util.concurrent.ScheduledThreadPoolExecutor$DelayedWorkQueue.take(ScheduledThreadPoolExecutor.java:809) at java.util.concurrent.ThreadPoolExecutor.getTask(ThreadPoolExecutor.java:1067) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1127) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) at java.lang.Thread.run(Thread.java:745) "pool-1-thread-19" #63 prio=5 os_prio=0 tid=0x0000000000a1e800 nid=0x1bff waiting on condition [0x00007fa109f56000] java.lang.Thread.State: WAITING (parking) at sun.misc.Unsafe.park(Native Method) - parking to wait for <0x00000000d3ab9e50> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject) at java.util.concurrent.locks.LockSupport.park(LockSupport.java:175) at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:2039) ... Linux thread ID in hex jstack forces STW pause in target JVM! 

Thread CPU usage in JVM sjk ttop command - https://github.com/aragozin/jvm-tools 2016-07-27T07:47:20.674-0400 Process summary process cpu=8.11% application cpu=2.17% (user=1.52% sys=0.65%) other: cpu=5.95% GC cpu=0.00% (young=0.00%, old=0.00%) heap allocation rate 1842kb/s safe point rate: 1.1 (events/s) avg. safe point pause: 0.43ms safe point sync time: 0.01% processing time: 0.04% (wallclock time) [003120] user= 1.12% sys= 0.24% alloc= 983kb/s - RMI TCP Connection(1)-172.17.168.11 [000039] user= 0.30% sys= 0.26% alloc= 701kb/s - DB feed - UserPermission.DBWatcher [000053] user= 0.00% sys= 0.05% alloc= 50kb/s - Statistics [000038] user= 0.00% sys= 0.05% alloc= 4584b/s – Reactor-0 [000049] user= 0.00% sys= 0.03% alloc= 38kb/s - DB feed - UserInfo.DBWatcher [000036] user= 0.00% sys= 0.03% alloc= 0b/s - Abandoned connection cleanup thread [003122] user= 0.00% sys= 0.03% alloc= 4915b/s - JMX server connection timeout 3122 [000040] user= 0.10% sys=-0.09% alloc= 8321b/s - DB feed - Report.DBWatcher [000050] user= 0.00% sys= 0.01% alloc= 24kb/s - DB feed - Rule.DBWatcher [000051] user= 0.00% sys= 0.01% alloc= 9034b/s - DB feed - EmailAccount.DBWatcher [000044] user= 0.00% sys= 0.01% alloc= 4840b/s - DB feed - Analytics.DBWatcher [000041] user= 0.00% sys= 0.01% alloc= 9999b/s - DB feed - Contact.DBWatcher [000054] user= 0.00% sys= 0.01% alloc= 3481b/s – Statistics [000001] user= 0.00% sys= 0.00% alloc= 0b/s - main [000002] user= 0.00% sys= 0.00% alloc= 0b/s - Reference Handler [000003] user= 0.00% sys= 0.00% alloc= 0b/s – Finalizer 

Thread CPU usage in JVM Mission Control (JDK tool) 

Java Threads - Conclusion Java threads are native OS threads  Use Linux diagnostic tools -XX:+PreserveFramePointer – make Java stack “walkable” JIT symbol generation - https://github.com/jvm-profiling-tools/perf-map-agent  Exploit taskset to control CPU affinity Control number of system Java threads  Limit number of parallel GC threads -XX:ParallelGCThredas 

Networking and IO 

Network tuning Cross region data transfers (client or server)  Tune options at socket level  Tune Linux network caps (sysctl) net.ipv4.tcp_rmem net.ipv4.tcp_wmem UDP based communications net.core.wmem_max net.core.rmem_max 

Leaking OS resources Linux OS has number cap on file handles if exceeded …  Cannot open new files  Cannot connect / accept socket connections 

Leaking OS resources Linux OS has number cap on file handles Java Garbage collector closes handles automatically  Files and sockets  Eventually … 

Leaking OS resources Linux OS has number cap on file handles Java Garbage collector closes handles automatically  Files and sockets  Eventually … Best practices  Always close your files and sockets explicitly  You should explicitly close socket object after SocketException 

Leaking OS resources Resources which cannot be explicitly disposed  File memory mappings  NIO direct buffers Diagnostics  Java heap dump can be analyzed for objects pending finalization 

Conclusion 

Summary You must size JVM  Heap size = young space + live set + reserve  JVM footprint = heap size + extra You can use native Linux diagnostic tools for JVM  Tip: you can use JDK tools with Linux core dump (requires debug symbols for OpenJDK) Linux tuning  Beware THP (Transparent Huge Pages)  Do network tuning on non-frontend servers too  Exploit NUMA and thread affinity 

Links Java Memory Tuning and Diagnostic  http://blog.ragozin.info/2016/10/hotspot-jvm-garbage-collection-options.html  https://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot/tooldescr007.html  Using JDK tools with Linux core dumps https://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot/bugreports004.html#CHDHDCJD Linux Transparent Huge Pages reading  https://www.perforce.com/blog/tales-field-taming-transparent-huge-pages-linux  https://tobert.github.io/tldr/cassandra-java-huge-pages.html  https://alexandrnikitin.github.io/blog/transparent-hugepages-measuring-the-performance-impact/ Profiling and performance monitoring  https://github.com/jvm-profiling-tools/perf-map-agent  https://github.com/aragozin/jvm-tools 

Thank you Alexey Ragozin [email protected] https://blog.ragozin.info https://github.com/aragozin