Let's talk about Garbage Collection

Garbage Collectors

Haim Yadid, Head of Performance and Application Infra Group

Motivation & Goals

Why Java has GC?

Memory management is hard malloc/free

Memory leaks (dangling pointers) Heap corruption (Free an object twice)

Reference counting Cyclic graphs

Garbage Collectors

GCMe allocate

myObject = new MyClass(2015) GC cleans So everything is cool !

Goal: Minimize Memory Overhead

Garbage collector memory overhead Internal structures Additional memory required for GC Policy and amount of memory generated before GC

Garbage Collectors

Goal: Application Throughput

No garbage collection at all --> Application throughput 1 (100% of the time) If in average garbage collection

consumes x milliseconds Every y milliseconds

Throughput is (y-x)/y E.g. if GC consumes 50 ms every second Throughput is 95% (GC overhead is 5%)

Garbage Collectors

GCtime ————— Total Time

Goal: Responsiveness

Pause Time Garbage collector stops the application During that time the application is not responsive What is the maximal delay your application can sustain:

Batch applications seconds… Web applications ½ second? Swing UI 100 ms max Trading application milliseconds Robotics microseconds….

Garbage Collectors

Goals

Memory footprint Throughput Max pause time

Garbage Collectors: G1

Contradicting! Choose 2 of 3

Goal: Fast Allocation

Maintaining free list of objects tend to lead to fragmentation Fragmentation increases allocation time A well known problem of C/C++ programs

Garbage Collectors

Goal: Locality

TLAB - thread local allocation buffer Maintaining locality utilizes CPU cache Linear allocation mechanism Per thread buffer allocated on first object alloc after new Gen GC. Small objects allocated linearly on that buffer FAST 10 machine instructions

Resized automatically ( ResizeTLAB=true) Based on TLABWasteTargetPercent=1

Garbage Collectors

Terminology

Sizing Heap

Heap the region of objects reside |H| - The size of the heap #H - number of object in the heap

Garbage Collectors

GC Root

An object which is references from outside the heap or heap region.

Java Local JNI Native stack System class Busy monitor Etc…

Garbage Collectors

Sizing : Live Set

Live Set : Object that are reachable from garbage collection roots #LS - number of object in the live set |LS| - The size of live set

Garbage Collectors

or heap region. Java Local JNI Native stack System class Busy monitor Etc…

Mutator

Application threads The stuff that create new objects and changes state The stuff that makes the garbage collector suffer Allocation rate - how fast new objects are allocated Mutation rate - how fast app changes references of live objects

Garbage Collectors

Collector

Pauseless Collector (Concurrent Collector )vs Stop the world (STW) Collector Serial Collector -Single threadedvs Parallel collector -Multi threaded Incremental vs Monolithic Conservative vs Precise

Garbage Collectors

GC Safepoint

A point in thread execution where it is safe for a GC to operate References inside thread stacks are identifiable If a thread is in a safe point GC may commence Thread will not be able to exit a safe-point until GC ends Global Safe-point : All threads enter a safe point Safe points must be frequent

Garbage Collectors

Building Blocks: Mark and Sweep

Mark: O(#LS) Every object holds a bit initially set to 0 Start with GC roots DFS traversal every visited object change to 1

Sweep O(|H|) Objects with 0 are added to the empty list

Downside: Heap fragmentation

Garbage Collectors

AB

C

E F

D A B C D F

Building Blocks: Copy CollectorMark: O(#LS)

Same as mark and sweep Copy: O(|LS|) = O(#LS)

All live objects are copied to an empty region(to space) No fragmentation Downside:

Need twice as much memory Copy is very expensive STW - mutators cannot work at the same time as copy

Garbage Collectors

A B C D F A B C D F

from space to space

Building Blocks: Mark (Sweep)Compact

Mark: O(#LS) Same as mark&sweep

Sliding compaction O(|H|+|LS|) move objects (relocate) fix pointers(remap)

Compact to the beginning of the heap Do not need twice the memory Copy is more delicate and may be slower

Garbage Collectors

The Weak Generational Hypothesis

Most objects survive for only a short period of time. Low number of references from old to new objects

Garbage Collectors

Generational Garbage Collectors

Since JDK 1.2 all collectors are generational and advantage of the WGH Different Collectors can be chosen for each generation

New generation collector Tenured generation collector

GC roots to young gen are maintained by a “remembered set”

Garbage Collectors

Oracle Hotspot Garbage Collectors

Serial (Serial, MSC) Train Collector (history) Parallel Collectors (a.k.a throughput collector) Concurrent collector (CMS) iCMS incremental CMS G1GC (Experimental)

Garbage Collectors

Major Memory Regions

Monitoring the JVM

Young PermTenured Code Cache

Young generation Further divided into:

Eden A “from” survivor space A “to” survivor space

Tenured (old) generation Permanent generation/Meta Space Code Cache

Heap Non Heap

NativeNative

Object Life Cycle

Most objects are allocated in Eden space. When Eden fills up a minor GC occurs Reachable object are copied “to” survivor space. There are two survivor spaces surviving objects are copied alternately from one to the other.

Eden S0S1

Eden S0S1

Eden S0S1

1 2 3

Object Promotion

Objects are promoted to the old generation (tenured) when:

surviving several minor GCs Survival spaces fill up

Serial Collectors

Single threaded Stop the world Monolithic Efficient (no communication between threads) The default on certain platforms

Garbage Collectors:Serial

Serial New Collector

-XX:+ UseSerialGC Serial new Single threaded young generation collector Triggered when:

Eden space is full. An explicit invocation or call to System.gc().

Garbage Collectors: Serial

MSC (Serial Old)

Single threaded tenured generation collector Mark & Sweep Compact Events which initiate a serial collector garbage collection

Tenured generation space is unable to satisfy an object promotion coming from young generation. An explicit invocation or call to System.gc().

Garbage Collectors: Serial

Serial Collector: Suitability

Well suited for single processor core machines CPU affinity (one-to-one JVM to processor core configuration) Tends to work well for applications with small Java heaps, i.e. less than 128mb- 256mb

Garbage Collectors:Serial

Train Collector

Introduced in Java 1.3 Divides the heap into small chunks Incremental Experimental Discontinued on Java 1.4

Garbage Collectors

Parallel Collector

Multi-threaded Monolithic Stop the world Three variants

-XX:+UseParallelGC -XX:+UseParallelOldGC

The default on most platforms

Garbage Collectors

Managing collector threads

Number parallel collector threads controlled by -XX:ParallelGCThreads=<N>

Defaults to Runtime.availableProcessors(). In a JDK 6 update release, 5/8ths available processors if > 8 Multiple JVM per machine configurations, set -XX: ParallelGCThreads=<N> such that sum of all threads < NCPU

Garbage Collectors

Parallel GC Triggering

Same as serial collector Events which initiate a minor garbage collection

Eden space is unable to satisfy an object allocation request. Results in a minor garbage collection event.

Events which might initiate a full garbage collection Tenured generation space unable to satisfy an object promotion coming from young generation. An explicit invocation or call to System.gc().

Garbage Collectors

Parallel Collector:Suitability

Reduce garbage collection overhead on multi-core processor systems Reduce pause time on multi-core systems Best throughput Pause time may be reasonable when heap size < 1GB

Garbage Collectors

CMS Collector

(Mostly) Concurrent mark and sweep tenured space collector Runs mostly concurrently with application threads Do not compact heap! Enabled with -XX:+UseConcMarkSweepGC ParNew - Parallel, multi-threaded young generation collector enabled by default. working with CMS

Garbage Collectors

Alas!

Lower throughput Requires more memory 20-30% more Concurrent mode failure will fallback to a stop the world full GC can occur when :

objects are copied to the tenured space faster than the concurrent collector can collect them. (“loses the race”) space fragmentation -XX:PrintFLSStatistics=1

Garbage Collectors

Concurrent Collector Phases

Concurrent collector cycle contains the following phases:

Initial mark (*) Concurrent mark Concurrent Pre-clean Remark (*) - second pass Concurrent sweep Concurrent reset

Garbage Collectors

The Concurrent Collector

Initial mark phase(*) Objects in the tenured generation are “marked” as reachable including those objects which may be reachable from young generation. Pause time is typically short in duration relative to minor collection pause times.

Concurrent mark phase Traverses the tenured generation object graph for reachable objects concurrently while Java application threads are executing.

Garbage Collectors

The Concurrent Collector Phases

Remark(*) Finds objects that were missed by the concurrent mark phase due to updates by Java application threads to objects after the concurrent collector had finished tracing that object.

Concurrent sweep Collects the objects identified as unreachable during marking phases.

Concurrent reset Prepares for next concurrent collection.

Garbage Collectors

PermGen Collection

Classes will not be collected during CMS concurrent phases Only during Full (STW) collection Explicitly instructed to do so using -XX:+CMSClassUnloadingEnabled and -XX:+PermGenSweepingEnabled, (the 2nd switch is not needed in post HotSpot 6.0u4 JVMs).

Garbage Collectors

Suitability

Application responsiveness is more important than application throughput More than one core Large Heaps > 1GB

Garbage Collectors

ExplicitGC and CMS

Explicit GC used to invoke the stop the world GC This can cause a problem with large heaps-XX:+ExplicitGCInvokesConcurrent (Java 6) -XX:+ExplicitGCInvokesConcurrentAndUnloadsClasses (requires Java6u4 or later).

Garbage Collectors

CMS Minor Collection Triggering

Minor collections are triggered as with the parallel collector The ParNew collector is built in such a way it can work in parallel with the CMS

Garbage Collectors

CMS Collection Triggering

Start if the occupancy exceeds a percentage threshold Default value is 92% -XX:CMSInitiatingOccupancyFraction=n where n is the % of the tenured space size.

Garbage Collectors

iCMS

Deprecated on Java 8 CMSIncrementalMode enables the concurrent modes to be done incrementally. Periodically gives additional processor back to the application resulting in better application responsiveness by doing the concurrent work in small chunks.

Garbage Collectors

Tune iCMS

CMSIncrementalMode has a duty cycle that controls the amount of work the concurrent collector is allowed to do before giving up the processor. Duty cycle is the % of time between minor collections the concurrent collector is allowed to run. Duty cycle by default is automatically computed using what's called automatic pacing. Both duty cycle and pacing can be fine tuned.

Garbage Collectors

Enabling iCMS Java 6

On JDK 6, recommend using the following two switches together:

-XX:+UseConcMarkSweepGC and -XX:+CMSIncrementalMode

Or use: -Xincgc

Garbage Collectors

Enabling iCMS Java5

On JDK 5 use all of the following switches together: -XX:+UseConcmarkSweepGC -XX:+CMSIncrementalMode -XX:+CMSIncrementalPacing -XX:CMSIncrementalDutyCycleMin=0 -XX:CMSIncrementalDutyCycle=10

JDK 5 settings mirror the default settings decided upon for JDK 6. JDK 5's -Xincgc!= CMSIncrementalMode, it enables CMS

Garbage Collectors

iCMS Fine Tuning

If full collections are still occurring, then: Increase the safety factor using -XX:CMSIncrementalSafetyFactor=n The default value is 10. Increasing safety factor adds conservatism when computing the duty cycle. Increase the minimum duty cycle using-XX:CMSIncrementalDutyCycleMin=n The default is 0 in JDK 6, 10 in JDK 5. Disable automatic pacing and use a fixed duty cycle using -XX:-CMSIncrementalPacing and -XX:CMSIncrementalDutyCycle=n The default duty cycle is 10 in JDK 6, 50 in JDK 5.

Garbage Collectors

G1 Garbage collector


Garbage First GC

Stop the world Incremental Beta stage in Java6 and Java 7 Supported since Java7u4


Garbage First GC

Apply New generation harvesting to the tenured Gen Achieve soft real time goal consume no more than x ms of any y ms time slice while maintaining high throughput for

programs with large heaps and high allocation rates, running on large multi-processor machines.


Heap Layout

divided into equal-sized heap regions, each a contiguous range of virtual memory Region size is 1-32MB based on heap size target 2000 regions A linked list of empty regions Heap is divided to New generation regions and old generation regions


Each region is either Marked Eden Survivor Space Old Generation Empty Humongous

Heap Layout


E E S SEE E E

O O O O OH H H

E E E

G1 New Gen GC

Live objects from young generation are moved to Survivor space regions Old gen regions

STW pause Calculate new size of eden and new Survivor space


G1 Concurrent Mark

Triggered when entire heap reaches certain threshold Mark regions Calculate liveliness information for each region Concurrent Empty regions are reclaimed immediately


OldGen collection

Choose regions with low liveliness Piggyback some during next young GC Denoted GC Pause (mixed)


Humongous Objects

1/2 of the heap region size allocated in dedicated (contiguous sequences of) heap regions; these regions contain only the humongous object GC is not optimized for these objects


Command line Options

-XX:+UseG1GC -XX:MaxGCPauseMillis=200 -XX:InitiatingHeapOccupancyPercent=45


Valid Combinations

Tenured


Young

G1GC

Parallel ScavengeParNewSerial

Serial OldCMS Parallel

Old

Summary: -XX flags for JDK6


Collector -XX: Param "Serial" + "Serial Old“ UseSerialGC "ParNew" + "Serial Old“ UseParNewGC "ParNew"+"CMS" + "Serial Old“ * UseConcMarkSweepGC

"Parallel Scavenge" + "Serial Old" UseParallelGC

"Parallel Scavenge" + "Parallel Old" UseParallelOldGC

G1 Garbage collector ** UseG1GC

Alternatives


Azul C4

A proprietary JVM Commercial Pause-less Requires changes in the OS (Linux only) Achieves Throughput and Pause time May require more memory…. 32GB and above Useful for large heaps (1TB is casual)

Garbage Collectors: C4

Shenandoah

JEP 189 for Open JDK Developed in RedHat An Ultra-Low-Pause-Time Garbage Regions same as G1 Concurrent collection w/o memory barrier Not a generational collector Brooks forwarding pointer

Garbage Collectors: Shenandoah

Ref

Object

Weird References


Object Life Cycle

Garbage Collectors: References

Created

Initialized Strongly Reachable

Softly Reachable

Weakly Reachable

Finalized

Phantom Reachable

Finalizers

Create performance issues For example: do not rely on a finalizer to close file descriptors. Try to limit use of finalizer as safety net Use other mechanisms for releasing resources. Keep the work being done as short as possible.


Finalizers and GC

Inside a finalizer you have a reference to your object and technically you may resurrect it. Objects which have a finalize method will need two cycles of GC in order to be collected Can lead to OOM errors A resurrected object may be reachable again but it finalize method will not run again. In order to prevent this problem use phantom references….


Finalizers and Thread Safety

Finalizers are executed from a special thread According to the Java memory modelupdates to local variable may not be visible to the Finalization thread Occurs when GC happen too soon In order to ensure correct memory visibility one need to use a sync block to force coherency


Reference Objects drawbacks

lots of reference objects also give the garbage collector more work to do since unreachable reference objects need to be discovered and queued during garbage collection.

Reference object processing can extend the time it takes to perform garbage collections, especially if there are consistently many unreachable reference objects to process.


GC Tuning

Tuning Memory: Tuning GC

Sizing Generations

VisualVM's Monitor tab Jconsole's Memory tab VisualGC's heap space sizes GCHisto jstat's GC options -verbose:gc heap space sizes -XX:+PrintGCDetails heap space sizes


Logging GC activity

The three most important parameters are: -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintGCDateStamps -Xloggc:<logfile>

If you want less data -verbose:gc


Log File Rotation

-XX:+UseGCLogFileRotation Control log file size

-XX:GCLogFileSize=1000 Number of log files

-XX:NumberOfGCLogFiles=10 Starting from Java 7


Verbose GC on runtime

Java.lang.Memory Set verbose attribute to true


The GC Log

Log every GC occurs PrintGCDetails (more verbose)

[GC [DefNew: 960K->64K(960K), 0.0047410 secs] 3950K->3478K(5056K), 0.0047900 secs]

Verbose:gc (less verbose) [GC 327680K->53714K(778240K), 0.2161340 secs]

Before->After(Total), time


CMS Collector

Minor collections follow serial collector format. [Full GC [CMS: 5994K->5992K(49152K),0.2584730 secs] 6834K->5992K(63936K), [CMS Perm: 10971K->10971K(18404K)], 0.2586030 secs]

[GC [1 CMS-initial-mark: 13991K(20288K)] 14103K(22400K), 0.0023781 secs] [CMS-concurrent-preclean: 0.044/0.064 secs] [GC [1 CMS-remark: 16090K(20288K)] 17242K(22400K), 0.0210460 secs]


Additional information

More information with the flags -XX:+PrintGCApplicationStoppedTime -XX:+PrintGCApplicationConcurrentTime -XX:+PrintTenuringDistribution


GC Log Quirks

CMS GC logs may get garbled Due to concurrency of the different GC mechanisms Analysis tools should be able to overcome this problem Not easy at all and may need to fix manually.


J9

-Xverbosegclog:<file path> Structured xml


<gc-end id="92" type="scavenge" contextid="88" durationms="4.464" usertimems="4.000" systemtimems="0.000" timestamp="2014-01-21T16:37:18.953"> <mem-info id="93" free="5472664" total="8388608" percent="65"> <mem type="nursery" free="655360" total="2097152" percent="31"> <mem type="allocate" free="655360" total="1179648" percent="55" /> <mem type="survivor" free="0" total="917504" percent="0" /> </mem> <mem type="tenure" free="4817304" total="6291456" percent="76"> <mem type="soa" free="4502936" total="5977088" percent="75" /> <mem type="loa" free="314368" total="314368" percent="100" /> </mem> <pending-finalizers system="2" default="0" reference="24" classloader="0" /> <remembered-set count="1770" /> </mem-info> </gc-end>

Visualization Tools


JClarity: Cesnum

Commercial product Give recommendations


GCViewer

Open source originally by Tagtraum industry A new fork exists in github latest version 1.33 https://github.com/chewiebug/GCViewer Supports G1GC Java6/7 Not bullet proof


GCViewer


GCViewer View

Choose which info to view


GCViewer -Summary


GCViewer - Memory


GCViewer -Pause


VisualGC


Explicit GC

Do not use System.gc() unless there is a specific use case or need to. Disable Explicit GC: -XX:+DisableExplicitGC. Default RMI distributed GC interval is once per minute, (60000 ms).

Use: -Dsun.rmi.dgc.client.gcInterval =3600000 -Dsun.rmi.dgc.server.gcInterval =3600000 When using JDK 6 and the Concurrent collector also use -XX:+ExplicitGCInvokesConcurrent


Interpretation


The Jigsaw effect

Garbage collection makes heap diagrams over time look like a jigsaw. Minor collections visualize as small teeth Full collections visualize as large teeth


Measuring Memory Usage

Best way is heap dump From GC logs look on the lower points of the Full GC lines.


60.0

90.0

120.0

150.0

180.0

Memory Leak

Draw the line between the full gc points Increasing over time --> memory leak


60.0

90.0

120.0

150.0

180.0

Low Throughput

Under 95% Increase Heap:

Low throughput is usually a result of insufficient memory GC kicks in too frequently and frees small amounts

New Gen too small New gen GC kicks in too fast Not able to release enough as a result mid term objects spill to old gen.

Old Gen too small Application state spills to new Gen. Breaks the Generational Hypothesis


Contradicting Goals

Goal1: Retain as many objects as possible in the survivor spaces

Less promotion into the old generation Less frequent old GCs

Goal2: Do not copy very long- lived objects between the survivors

Unnecessary overhead on minor GCs


High Pause Time

Choose The correct GC scheme: CMS/G1 when low pause is required

Reduce footprint of user interactive processes Reduce memory allocations do not allocated too many related temporary objects chunking of work.


Software

Let's talk about Garbage Collection