JVM Garbage Collection Tuning Presentation Material. An excerpt from the Oracle's official document.

1. JVM Garbage
Collection Tuning
Heejong Lee
Reference: JVM Garbage Collection Tuning (http://www.oracle.com/technetwork/java/javase/gc-tuning-6-140523.html)

2. Pause-the-world garbage collector’s performance loss in parallelized systems

3. Lifetime distribution of objects
Efficient collection is made possible by focusing on the fact that
a majority of objects die young

4. Lifetime of objects
Eden
Survivor 1
Survivor 2
Tenured

5. • The percentage of total time
not spent in garbage
collection
• Includes time spent in
allocation
Two Primary Measures of Garbage Collection
Performance
• Times when an application
appears unresponsive
because garbage collection is
occurring
Throughput Pauses

6. • The working set of one
process, measured in pages
and cache lines
• May dictate scalability
Two Alternative Measures of Garbage Collection
Performance
• The time between when an
object becomes dead and
when the memory becomes
available
Footprint Promptness

7. Trade-off between Measures of Garbage Collection
Performance
• A very large young generation space may maximize throughput,
but does so at the expense of footprint, promptness and pause
times
• A small young generation space can minimize young generation
pauses at the expense of throughput

8. • Metrics particular to the
application
• ex) client load generator for
web servers, pmap for
daemons
Measurement of Garbage Collection Performance
• The diagnostic output of
virtual machine itself
Throughput, Footprint Pauses

9. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]

10. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]
minor collection (young generation collection)

11. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]
major collection (all generation collection)

12. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]
the combined size of live objects before and after gc

13. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]
the amount of space usable for java objects without requesting more memory from the OS
(except permanent generation and one of the survivor spaces)

14. Measurement of Garbage Collection Performance
• `-verbose:gc` causes information about the heap and garbage
collection
Pauses
[GC 325407K->83000K(776768K), 0.2300771 secs]
[GC 325816K->83372K(776768K), 0.2454258 secs]
[Full GC 267628K->83769K(776768K), 1.8479984 secs]
the time taken to perform the collection

15. Measurement of Garbage Collection Performance
• `-XX:+PrintGCDetails` causes additional information about the
heap and garbage collection
Pauses
[GC [DefNew: 64575K->959K(64576K), 0.0457646 secs] 196016K->133633K(261184K), 0.0459067 secs]
the log from young generation

16. Measurement of Garbage Collection Performance
• `-XX:+PrintGCDetails` causes additional information about the
heap and garbage collection
Pauses
[GC [DefNew: 64575K->959K(64576K), 0.0457646 secs] 196016K->133633K(261184K), 0.0459067 secs]
the log from entire heap

17. Measurement of Garbage Collection Performance
• `-XX:+PrintGCTimeStamps` will add a time stamp at the start of
each collection
Pauses
111.042: [GC 111.042: [DefNew: 8128K->8128K(8128K), 0.0000505 secs]111.042: [Tenured: 18154K-
>2311K(24576K), 0.1290354 secs] 26282K->2311K(32704K), 0.1293306 secs]

18. The Difference between Memory Size Ratios

19. Sizing the Generations
Total Heap
• `-XX:MinHeapFreeRatio=` : the generation will be expanded to
maintain the given ratio (in percent) of free space
• `-XX:MaxHeapFreeRatio=` : the generation will be contracted to
maintain the given ratio (in percent) of free space
• `-Xms` : lower bound of heap size
• `-Xmx` : upper bound of heap size

20. Sizing the Generations
The Rules of Adjusting Total Heap
• Try granting as much memory as possible unless you have
problems with pauses
• Setting `-Xms` and `-Xmx` to the same value increases
predictability by removing the most important sizing decision,
however the virtual machine cannot compensate a poor choice

21. Sizing the Generations
Young Generation
• `-XX:NewRatio=` : the ratio between young and tenured
generations. For example, 3 means the combined size of the eden
and survivor spaces will be one fourth of the total heap size
• `-XX:NewSize=` : lower bound of the young generation size
• `-XX:MaxNewSize=` : upper bound of the young generation size

22. Sizing the Generations
Young Generation
• The bigger the young generation, the less often minor collection
occur
• However, (because the total heap size is bounded) which will
increase the frequency of major collection

23. Sizing the Generations
Young Generation (Survivor Space)
• `-XX:SurvivorRatio=` : the ratio between eden and survivor space.
For example, 6 means each survivor space will be one sixth the
size of eden
• `-XX:+PrintTenuringDistribution` : show tenuring threshold and the
ages of objects in the young generation

24. Sizing the Generations
Young Generation (Survivor Space)
• If survivor spaces are too small, copying collection overflows
directly into the tenured generation
• The threshold controlled the number of times an object can be
copied before it is tenured is chosen to keep the survivor half full

25. Garbage Collectors
Serial Collector
• Uses single thread
• Relatively efficient since there is no communication overhead
between threads
• `-XX:+UseSerialGC`
Parallel Collector
(Throughput Collector)
• Performs minor collections in parallel (`-XX:+UseParallelGC`)
• Performs major and minor collections in parallel (`-XX:
+UseParallelOldGC`, available since 5u6, default since 7u4)
• Intended for applications with medium- to large-sized data
sets that are run on multi-threaded hardware
Concurrent Collector
• Performs most of its work concurrently
• Intended for applications with medium- to large-sized data
sets for which response time is more important than overall
throughput
• `-XX:+UseConcMarkSweepGC`

26. Selecting a Collector
Application has a small data set (up to
100MB)
• `-XX:+UseSerialGC`
Application will be run on a single
processor & no pause time requirement
• Let VM select the collector
• Or, `-XX:+UseSerialGC`
Peak application performance is the first
priority & no pause time requirement
• Let VM select the collector
• Or, `-XX:+UseParallelGC`, `-XX:+UseParallelOldGC`
Response time is more important than
overall throughput & pause time must be
kept short
• `-XX:+UseConcMarkSweepGC`

27. Parallel Collector
• On default, uses N garbage collector threads on a machine with N
processors
• One processor (serial collector > parallel collector)
• Two processors (serial collector < parallel collector)
• Three or more processors ( serial collector <<< parallel collector)
• `-XX:ParallelGCThreads=` controls the number of garbage
collector threads

28. Parallel Collector
• On default, uses N garbage collector threads on a machine with N
processors
• One processor (serial collector > parallel collector)
• Two processors (serial collector < parallel collector)
• Three or more processors ( serial collector <<< parallel collector)
• `-XX:ParallelGCThreads=` controls the number of garbage
collector threads

29. Ergonomics
• The parallel collector is selected by default on server-class machine
(since J2SE5)
• Automatic tuning is available by three goals (order by priority):
• Maximum garbage collection pause time (-XX:MaxGCPauseMillis=)
• Throughput (-XX:GCTimeRatio=)
default value is 99, resulting in a goal of 1% of the time in GC
• Footprint (-Xmx)

30. Default Heap Size
• Initial heap size : memory / DefaultInitialRAMFraction
(default: memory / 64)
• Maximum heap size : MIN(memory / DefaultMaxRAMFraction, 1GB)
(default: MIN(memory / 4), 1GB)

31. OutOfMemoryError
• If more than 98% of the total time is spent in garbage collection and
less than 2% of the heap is recovered
• Can be disabled by `-XX:-UseGCOverheadLimit`

32. Concurrent Collector
• Reduces the time required for major collection (low pause time): suitable
for applications which have a relatively large set of long-lived data and
run on machines with two or more processors
• Only two short pause during each major collection (mark and remark)
• Most of the work in the major collection (tracing of live objects, sweeping
of unreachable objects) is done concurrently with the application
• Minor collections are done in a manner similar to the parallel collector
(the application threads are stopped during the collection)
• Concurrent mode failure when GC fails to completing the collection
before either the tenured and permanent generation becomes full

33. OutOfMemoryError (of Concurrent Collection)
• If more than 98% of the total time is spent in garbage collection and
less than 2% of the heap is recovered
• Can be disabled by `-XX:-UseGCOverheadLimit`
• Only collections performed while the application is stopped count
toward excessive GC time (typically due to a concurrent mode failure
or an explicit call to System.gc())

34. Floating Garbage (of Concurrent Collection)
• Objects that became unreachable by the time collection finishes and
will be collected during the next collection cycle
• A rough rule of thumb: try increasing the size of the tenured generation
by 20% to account for the floating garbage

35. Concurrent Phases
Mark
Tracing reachable
object graph
Remark
Sweeping
unreachable objects Wait
Pause Consume CPU resource Pause Consume CPU resource
Use minimum
resource

36. Starting Concurrent Collection Cycle
• The concurrent collector maintains estimate time remaining before the
tenured generation will be exhausted and of the time needed for a
concurrent collection cycle
• A concurrent collection cycle will be started based on these dynamic
estimates
• Will also start if the occupancy of the tenured generation exceeds an
initiating occupancy. This percentage value can be adjusted with the
option `-XX:CMSInitiatingOccupancyFraction=`

37. Incremental Mode
• Breaking up the concurrent phases into short burst of activity, which
are scheduled to occur mid-way between minor pauses
• Especially useful when applications that need the low pause times are
run on machines with small numbers of processors (e.g., 1 or 2)

38. Incremental Mode
• `-XX:+CMSIncrementalMode`: enables incremental mode. must also
be enabled with `-XX:+UseConcMarkSweepGC`
• `-XX:+CMSIncrementalPacing`: enables automatic pacing. The
incremental mode duty cycle is automatically adjusted based on
statistics collected while the JVM is running
• `-XX:CMSIncrementalDutyCycle=`: the percentage of time between
minor collections that the concurrent collector is allowed to run. Just
initial value when CMSIncrementalPacing is enabled

39. Recommended Option for Incremental Mode
-XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode
-XX:+PrintGCDetails -XX:+PrintGCTimeStamps
-XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode
-XX:+PrintGCDetails -XX:+PrintGCTimeStamps
-XX:+CMSIncrementalPacing -XX:CMSIncrementalDutyCycleMin=0
-XX:CMSIncrementalDutyCycle=10
Java 6
Java 5

40. Explicit Garbage Collection
• `-XX:+DisableExplicitGC` causes the VM to ignore calls to System.gc()