Java Garbage Collector:
Friend or Foe
Krasimir Semerdzhiev
Development Architect / SAP Labs Bulgaria
Agenda
1. Brief historical view
2. Myths and Urban legends
3. GC machinery
4. Try to stay out of trouble
History of Garbage Collection
A long time ago, in a galaxy far far away…
* Counting from [McCarthy 1959]
McCarthy (1959)
LISt Processor (LISP)
Reference counting (IBM)
Naive Mark/sweep GC
Semi-space collector
1959*
Knuth (1973/1978)
Copying collector
Mark/sweep GC
Mark/don’t sweep GC
1970 1990 2000 2011
Appel (1988), Baker (1992)
Generational GC
Train GC
Stop the world
Cheng-Blelloch (2001)
Concurrent
Parallel
Real-Time GC
2003
Bacon, Cheng, Rajan (2003)
Metronome GC
Garbage first GC (G1)
Ergonomics
20091995
Agenda
1. Brief historical view
2. Myths and Urban legends
3. GC machinery
4. Try to stay out of trouble
5. Tools for the masses
Myths
Java and C/C++ performance
Is C/C++ faster than Java?
The short answer: it depends.
/Cliff Click
Myths
My GC cleans up tons of memory – what’s going on.
String s = "c"?
= ref + 8 + (12 + 2) + 4 + 4 + 4 > 34 bytes.
/** The value is used for character storage. */
private final char value[];
/** The offset is the first index of the storage that is used. */
private final int offset;
/** The count is the number of characters in the String. */
private final int count;
/** Cache the hash code for the string */
private int hash; // Default to 0
/** use serialVersionUID from JDK 1.0.2 for interoperability */
private static final long serialVersionUID =
-6849794470754667710L;
Is Java bad at memory management?
public static void main(String[] args) throws Exception {
final long start = Runtime.getRuntime().freeMemory();
final byte[][] arrays = new byte[100][];
for (int i = 0; i < arrays.length; i++) {
arrays[i] = new byte[100];
long current = Runtime.getRuntime().freeMemory();
System.out.println(start + " " + current);
Thread.sleep(1000);
}
}
Agenda
1. Brief historical view
2. Myths and Urban legends
3. GC machinery*
4. Try to stay out of trouble
5. Tony Printezis
* Credits for the GC insights go to
Tony Printezis and his excellent
J1 sessions
Garbage collector
Let’s start with a simple memory example…
Garbage collector
Mark-Sweep
Root object
Mark-Sweep
Garbage collector
Mark-Sweep
Root object
Mark-Sweep – Marking…
Garbage collector
Mark-Sweep
Root object
Free List Root
Mark-Sweep – Sweeping…
Garbage collector
Let’s try another one …
Garbage collector
Mark-Compact
Root object
Mark-Compact
Garbage collector
Mark-Compact
Root object
Mark-Compact – Marking…
Garbage collector
Mark-Compact
Root object
Mark-Sweep – Compacting…
Free Pointer
Garbage collector
Let’s try another one …
Garbage collector
Copying
Root object
Copying
From space
To space Free and unused
Garbage collector
Copying
Root object
Copying – Evacuation
From space
To space Free and unused
Garbage collector
Copying
Root object
Copying – Flipping
From space
To space Free and unused
Free Pointer
Garbage collector
Let’s try another one … ;o)
Garbage collector
Generational Garbage Collection
Generational Garbage Collection – moving to more modern times
Young generation
Old Generation
Allocations
Promotion
Java Heap
Hotspot JVM
Memory layout
Hotspot JVM (Java) heap layout
Young generation
Old Generation
Perm Generation
Everything
else
Java Heap
Hotspot JVM
Memory layout
Hotspot JVM (Java) heap layout
Young generation
Old Generation
Perm Generation
Everything
else
Maximum size is limited:
■ 32 bit -> 2Gb
■ 64 bit -> way larger
If 2Gb is the max for the process – you can’t get it all
for the Java heap only!
Hotspot JVM
Memory layout
Hotspot JVM (Java) heap layout
Survivor spaces
Old Generation
Perm Generation
Young generation
unused
Eden
From To
Hotspot JVM
Memory layout
Hotspot JVM (Java) – (Small) GC running
Survivor spaces
Old Generation
Perm Generation
Young generation
unused
Eden
From To
Hotspot JVM
Memory layout
Hotspot JVM (Java) – (Minor) GC running
Survivor spaces
Old Generation
Perm Generation
Young generation
unused
Eden
From To
Hotspot JVM
Memory details
-Xmx, -Xms, -Xmn
 Control the Java Object heap size only
 Doesn’t have impact on Perm size, native heap and the Thread stack size
-XX:PermSize, -XX:MaxPermSize
 Stores class definitions, methods, statis fields
 Common reason for OOM errors.
-Xss
 Configures the stack size of every thread.
-XX:+UseTLAB, -XX:-UseTLAB, -XX:+PrintTLAB
 Enables the Thread Local Allocation Buffer.
 Since Java SE 1.5 – this is automatically tuned to each and every thread.
TCP Connection buffer sizes – allocated in native space
 Use Socket.setSendBufferSize(int) and Socket.setReceiveBufferSize(int).
 OS will use the smaller of the two or will simply ignore that setting.
Object allocation statistics:
■ Up to 98% of new objects are
short-lived
■ Up to 98% die before another
Mb is allocated
Agenda
1. Brief historical view
2. Myths and Urban legends
3. GC machinery
4. Try to stay out of trouble
OutOfMemoryError
How to proceed?
java.lang.OutOfMemoryError: PermGen space
 Increase the Perm Space – will help if there is no code generation happening
 In case of a leak – the only solution is frequent system restart.
java.lang.OutOfMemoryError: unable to create new native thread
 Decrease –Xmx or –Xss
java.lang.StackOverflowError
 Increase –Xss or fix the corresponding algorithm
IOException: Too many open files (for example)
 Increase the OS file handle limit per process
 Check for leaking file handles
 Physical limit of the VM is 2048 ZIP/JAR files opened at the same time.
java.lang.OutOfMemoryError: Direct buffer memory
 Direct mapped memory – used for java.nio.ByteBuffer
 Increase -XX:MaxDirectMemorySize
Finalizer methods
Good or bad
Again short answer: it depends ;-)
Infrastructure components
 Might be used for debugging/tracing purposes
 Major scenario – closing of critical backend resources
 All finalizer methods are collected separately. No mass-wipe is performed on them!
 Never, ever throw an exception in a finalizer!
Applications
 Avoid finalizers by all means!
Per-request created objects
 Avoid finalizers by all means!
Finalizer queue
Working threads
Finalizer thread (singleton)
Response time peaks
without CMS
All those are full GCs…
No more full GCs…
Response time peaks
with CMS
Garbage Collection Strategies
Does it pay off to play with that?
-XX:+UseSerialGC
 Default state before Java 5. Obsolete!
-XX:+UseParallelGC - Parallel Scavange GC (1.4.2)
 Works on Young Generation only
 Use -XX:ParallelGCThreads to configure it
-XX:+UseParNewGC - Parallel New-Gen GC (5.0)
 Use also –XX:SurvivorRatio and –XX:MaxTenuringThreshold to
define the lifespan of objects in the eden space.
 -XX:+CMSClassUnloadingEnabled to trigger concurrent cleanup of
the Perm space
-XX:+UseConcMarkSweepGC - Concurrent Old-Gen GC
 Default since Java SE 5.
-XX:+UseParallelOldGC - Concurrent Old-Gen GC (5.0)
 Parallel Compaction of Old space
Terms:
• Serial – 1 thing at a
time
• Parallel – work is
done in multiple
threads
• Concurrent – work
happens
simultaneously
Garbage Collection Ergonomics
What’s that?
New way to configure GC – specify:
 Max pause time goal (-XX:MaxGCPauseMillis)
 Throughput goal (-XX:GCTimeRatio)
 Assumes minimal footprint goal
 Use -XX:+UseParallelGC with those.
 Garbage First (G1) GC will be the default in Java SE 7. Released
with JDK 1.6.0_u14 for non-productive usage.
 Enable that for testing via:
 -XX:+UnlockExperimentalVMOptions
 -XX:+UseG1GC
Analyzing Garbage Collection output
Getting GC output is critical for further analysis!
 -verbose:gc get 1 line per GC run with some basic inf
 -XX:+PrintGCDetails get more extended info (perm, eden, tenured)
 -XX:+PrintGCTimeStamps get timestamps since the start of the VM. Allows correlation
 -XX:-TraceClassUnloading get also the unloaded classes – helps tracing leaks
 -Xloggc:gc.log direct GC output to a special file instead of the process output
 -XX:+HeapDumpOnOutOfMemoryError Heap dump generation on OutOfMemory
GC Viewer
GC output viewer
 Import the gc.out file.
 Correlate over time
 So far the most
comprehensive viewer
 Stay tuned and monitor
the Eclipse news ;-)
Visual VM
Supplied by Oracle with JDK
 Free and very comprehensive
 Evolves together with the VM.
 Focus shifting from that to
Mission Control (the Jrockit
profiling solution)
Eclipse Memory Analyzer
Developed by SAP and IBM in Eclipse
 Track GC roots
 Do what-if analysis
 SQL-like query language
 Custom filters
 Track Leaking
classloaders
Garbage Collection – the universal settings
There are NO universal settings! Sorry. :(
 G1 is the first GC, trying to go in that direction and leave the self-tuning to the VM
 You have to test with realistic load!
 You have to test on realistic hardware!
 Tune the GC as you fix the memory leaks which will inevitably show up.
 Try to find the balance of uptime/restart intervals.
Contact Questions?
Krasimir Semerdzhiev
krasimir.semerdzhiev@sap.com

[BGOUG] Java GC - Friend or Foe

  • 1.
    Java Garbage Collector: Friendor Foe Krasimir Semerdzhiev Development Architect / SAP Labs Bulgaria
  • 2.
    Agenda 1. Brief historicalview 2. Myths and Urban legends 3. GC machinery 4. Try to stay out of trouble
  • 3.
    History of GarbageCollection A long time ago, in a galaxy far far away… * Counting from [McCarthy 1959] McCarthy (1959) LISt Processor (LISP) Reference counting (IBM) Naive Mark/sweep GC Semi-space collector 1959* Knuth (1973/1978) Copying collector Mark/sweep GC Mark/don’t sweep GC 1970 1990 2000 2011 Appel (1988), Baker (1992) Generational GC Train GC Stop the world Cheng-Blelloch (2001) Concurrent Parallel Real-Time GC 2003 Bacon, Cheng, Rajan (2003) Metronome GC Garbage first GC (G1) Ergonomics 20091995
  • 4.
    Agenda 1. Brief historicalview 2. Myths and Urban legends 3. GC machinery 4. Try to stay out of trouble 5. Tools for the masses
  • 5.
    Myths Java and C/C++performance Is C/C++ faster than Java? The short answer: it depends. /Cliff Click
  • 6.
    Myths My GC cleansup tons of memory – what’s going on. String s = "c"? = ref + 8 + (12 + 2) + 4 + 4 + 4 > 34 bytes. /** The value is used for character storage. */ private final char value[]; /** The offset is the first index of the storage that is used. */ private final int offset; /** The count is the number of characters in the String. */ private final int count; /** Cache the hash code for the string */ private int hash; // Default to 0 /** use serialVersionUID from JDK 1.0.2 for interoperability */ private static final long serialVersionUID = -6849794470754667710L;
  • 7.
    Is Java badat memory management? public static void main(String[] args) throws Exception { final long start = Runtime.getRuntime().freeMemory(); final byte[][] arrays = new byte[100][]; for (int i = 0; i < arrays.length; i++) { arrays[i] = new byte[100]; long current = Runtime.getRuntime().freeMemory(); System.out.println(start + " " + current); Thread.sleep(1000); } }
  • 8.
    Agenda 1. Brief historicalview 2. Myths and Urban legends 3. GC machinery* 4. Try to stay out of trouble 5. Tony Printezis * Credits for the GC insights go to Tony Printezis and his excellent J1 sessions
  • 9.
    Garbage collector Let’s startwith a simple memory example…
  • 10.
  • 11.
  • 12.
    Garbage collector Mark-Sweep Root object FreeList Root Mark-Sweep – Sweeping…
  • 13.
  • 14.
  • 15.
  • 16.
  • 17.
  • 18.
  • 19.
    Garbage collector Copying Root object Copying– Evacuation From space To space Free and unused
  • 20.
    Garbage collector Copying Root object Copying– Flipping From space To space Free and unused Free Pointer
  • 21.
    Garbage collector Let’s tryanother one … ;o)
  • 22.
    Garbage collector Generational GarbageCollection Generational Garbage Collection – moving to more modern times Young generation Old Generation Allocations Promotion
  • 23.
    Java Heap Hotspot JVM Memorylayout Hotspot JVM (Java) heap layout Young generation Old Generation Perm Generation Everything else
  • 24.
    Java Heap Hotspot JVM Memorylayout Hotspot JVM (Java) heap layout Young generation Old Generation Perm Generation Everything else Maximum size is limited: ■ 32 bit -> 2Gb ■ 64 bit -> way larger If 2Gb is the max for the process – you can’t get it all for the Java heap only!
  • 25.
    Hotspot JVM Memory layout HotspotJVM (Java) heap layout Survivor spaces Old Generation Perm Generation Young generation unused Eden From To
  • 26.
    Hotspot JVM Memory layout HotspotJVM (Java) – (Small) GC running Survivor spaces Old Generation Perm Generation Young generation unused Eden From To
  • 27.
    Hotspot JVM Memory layout HotspotJVM (Java) – (Minor) GC running Survivor spaces Old Generation Perm Generation Young generation unused Eden From To
  • 28.
    Hotspot JVM Memory details -Xmx,-Xms, -Xmn  Control the Java Object heap size only  Doesn’t have impact on Perm size, native heap and the Thread stack size -XX:PermSize, -XX:MaxPermSize  Stores class definitions, methods, statis fields  Common reason for OOM errors. -Xss  Configures the stack size of every thread. -XX:+UseTLAB, -XX:-UseTLAB, -XX:+PrintTLAB  Enables the Thread Local Allocation Buffer.  Since Java SE 1.5 – this is automatically tuned to each and every thread. TCP Connection buffer sizes – allocated in native space  Use Socket.setSendBufferSize(int) and Socket.setReceiveBufferSize(int).  OS will use the smaller of the two or will simply ignore that setting. Object allocation statistics: ■ Up to 98% of new objects are short-lived ■ Up to 98% die before another Mb is allocated
  • 29.
    Agenda 1. Brief historicalview 2. Myths and Urban legends 3. GC machinery 4. Try to stay out of trouble
  • 30.
    OutOfMemoryError How to proceed? java.lang.OutOfMemoryError:PermGen space  Increase the Perm Space – will help if there is no code generation happening  In case of a leak – the only solution is frequent system restart. java.lang.OutOfMemoryError: unable to create new native thread  Decrease –Xmx or –Xss java.lang.StackOverflowError  Increase –Xss or fix the corresponding algorithm IOException: Too many open files (for example)  Increase the OS file handle limit per process  Check for leaking file handles  Physical limit of the VM is 2048 ZIP/JAR files opened at the same time. java.lang.OutOfMemoryError: Direct buffer memory  Direct mapped memory – used for java.nio.ByteBuffer  Increase -XX:MaxDirectMemorySize
  • 31.
    Finalizer methods Good orbad Again short answer: it depends ;-) Infrastructure components  Might be used for debugging/tracing purposes  Major scenario – closing of critical backend resources  All finalizer methods are collected separately. No mass-wipe is performed on them!  Never, ever throw an exception in a finalizer! Applications  Avoid finalizers by all means! Per-request created objects  Avoid finalizers by all means! Finalizer queue Working threads Finalizer thread (singleton)
  • 32.
    Response time peaks withoutCMS All those are full GCs…
  • 33.
    No more fullGCs… Response time peaks with CMS
  • 34.
    Garbage Collection Strategies Doesit pay off to play with that? -XX:+UseSerialGC  Default state before Java 5. Obsolete! -XX:+UseParallelGC - Parallel Scavange GC (1.4.2)  Works on Young Generation only  Use -XX:ParallelGCThreads to configure it -XX:+UseParNewGC - Parallel New-Gen GC (5.0)  Use also –XX:SurvivorRatio and –XX:MaxTenuringThreshold to define the lifespan of objects in the eden space.  -XX:+CMSClassUnloadingEnabled to trigger concurrent cleanup of the Perm space -XX:+UseConcMarkSweepGC - Concurrent Old-Gen GC  Default since Java SE 5. -XX:+UseParallelOldGC - Concurrent Old-Gen GC (5.0)  Parallel Compaction of Old space Terms: • Serial – 1 thing at a time • Parallel – work is done in multiple threads • Concurrent – work happens simultaneously
  • 35.
    Garbage Collection Ergonomics What’sthat? New way to configure GC – specify:  Max pause time goal (-XX:MaxGCPauseMillis)  Throughput goal (-XX:GCTimeRatio)  Assumes minimal footprint goal  Use -XX:+UseParallelGC with those.  Garbage First (G1) GC will be the default in Java SE 7. Released with JDK 1.6.0_u14 for non-productive usage.  Enable that for testing via:  -XX:+UnlockExperimentalVMOptions  -XX:+UseG1GC
  • 36.
    Analyzing Garbage Collectionoutput Getting GC output is critical for further analysis!  -verbose:gc get 1 line per GC run with some basic inf  -XX:+PrintGCDetails get more extended info (perm, eden, tenured)  -XX:+PrintGCTimeStamps get timestamps since the start of the VM. Allows correlation  -XX:-TraceClassUnloading get also the unloaded classes – helps tracing leaks  -Xloggc:gc.log direct GC output to a special file instead of the process output  -XX:+HeapDumpOnOutOfMemoryError Heap dump generation on OutOfMemory
  • 37.
    GC Viewer GC outputviewer  Import the gc.out file.  Correlate over time  So far the most comprehensive viewer  Stay tuned and monitor the Eclipse news ;-)
  • 38.
    Visual VM Supplied byOracle with JDK  Free and very comprehensive  Evolves together with the VM.  Focus shifting from that to Mission Control (the Jrockit profiling solution)
  • 39.
    Eclipse Memory Analyzer Developedby SAP and IBM in Eclipse  Track GC roots  Do what-if analysis  SQL-like query language  Custom filters  Track Leaking classloaders
  • 40.
    Garbage Collection –the universal settings There are NO universal settings! Sorry. :(  G1 is the first GC, trying to go in that direction and leave the self-tuning to the VM  You have to test with realistic load!  You have to test on realistic hardware!  Tune the GC as you fix the memory leaks which will inevitably show up.  Try to find the balance of uptime/restart intervals.
  • 41.