Gargbage Collection (GC)

Gargbage Collection (GC)

Schemes
Schemes
Sang Shin
Sang Shin
Technology Evangelist Technology Evangelist
Sun Microsystems, Inc. Sun Microsystems, Inc.
www.javapassion.com www.javapassion.com
1
2
Agenda
Why you care on GC (as a eveloper!"
What is an Why Generational GC"
GC #er$ormance Metrics
GC %lgorithms
Types o$ Collectors
>
Serial collector
>
#arallel collector
>
#arallel compact collector
>
Concurrent Mar&'Sweep (CMS! collector( regular,
incremental
Ergonomics
3
Why you care on GC?
In general, the e$ault setting shoul wor& $ine $or most
applications
)or GC sensitive applications, however, choosing a
wrong GC scheme coul result in a less than esira*le
per$ormance
What is and Why
What is and Why
Generational GC?
Generational GC?
5
Issues of Non-Generational GC
Most straight$orwar GC will just iterate over every
o*ject in the heap an etermine i$ any other
o*jects re$erence it
>
This is the *ehavior o$ +on'generational GC
>
This gets really slow as the num*er o$ o*jects in the
heap increases
>
This oes not ta&e avantage o$ the characteristics o$
typical o*jects
,ence the reason $or Generational GC
6
Behavior of Typical Java O!ects
Typical o*ject (-oung o*ject! is most li&ely to ie
shortly a$ter it was create
>
It is calle .high in$ant mortality/
>
E0ample( 1ocal o*jects
2*jects that have *een aroun $or a while (2l
o*jects! will li&ely stay aroun $or a while
>
E0ample( 2*jects initiali3e at the time o$ application
startup
2nly a $ew re$erences $rom ol o*jects to young
o*jects e0ist
7
"#pirical $tatistics
Most o*jects are very short live
>
45'647 o$ all newly allocate o*jects
ie shortly a$ter they are create
>
45'647 o$ all newly allocate o*jects
ie *e$ore another mega*yte has *een allocate
8
%eap $pace With Generations
,eap space is organi3e into .generations/
>
-oung generation ($or young o*jects!
>
2l (or Tenure! generation ($or ol o*jects!
>
#erm generation (meta ata, classes, etc.!
&oung Generation
'er#anent Generation
Old Generation
9
'rinciples of Generational GC
8eep young an ol o*jects separate
>
Mi0ing them in a single space woul result in ine$$icient
GC
9se i$$erent GC algorithms $or i$$erent generation
heap spaces
>
2*jects in each generation have i$$erent li$e'cycle
tenencies
>
.9se the right tool $or the jo*/
10
Characteristics of &oung Generation
%eap $pace
GC:s occur relatively $re;uently
GC:s are e$$icient an $ast *ecause young
generation space is usually small an li&ely to
contain a lot o$ o*jects that are no longer re$erence
2*jects that survive some num*er o$ young
generation collections are promote, or tenure, to
ol generation heap space
11
Characteristics of Old Generational
%eap $pace
Typically larger than young generation heap space
Its occupancy grows more slowly
GC:s are in$re;uent *ut ta&es signi$icantly longer
time to complete
12
Generational GC
13
&oung Generation $pace (ayout
Mae o$ an area calle Eden plus two smaller survival
spaces
Most o*jects are initially allocate in Een
2ne o$ the two survival spaces hol o*jects that survive at
least one young generation GC while the other is empty
14
GC Algorith#s )sed
-oung generation
>
%lgorithms use emphasi3e spee since young
generation GC:s are $re;uent
2l generation
>
%lgorithms use emphasi3e e$$icient space since ol
generation ta&es up most o$ the heap space an have to
wor& with low gar*age ensities
When *oes GC
When *oes GC
Occur?
Occur?
16
When GC Occurs
When the young generation $ills up, a young
generation GC occurs
>
-oung generation GC is calle minor GC
When the ol generation oes not have enough
space $or the o*jects that are *eing promote, a )ull
GC occurs
>
)ull GC is also calle major GC
GC 'erfor#ance
GC 'erfor#ance
+etric
+etric
18
I#portant GC 'erfor#ance +etric
Throughput
>
The percentage o$ total time not spent in gar*age
collection, consiere over long perios o$ time.
#ause time
>
The length o$ time uring which application e0ecution is
stoppe while gar*age collection is *eing per$orme
19
Application ,e-uire#ent
<i$$erent applications have i$$erent re;uirements
>
,igher throughput is more important $or We* application(
pauses uring gar*age collection may *e tolera*le, or
simply o*scure *y networ& latencies.
>
Shorter pause time is more important to an interactive
graphics application
*e#o.
*e#o.
Collecting GC
Collecting GC
Infor#ation through
Infor#ation through
/isualGC0 !stat
/isualGC0 !stat
GC Algorith#s
GC Algorith#s
22
Choices of GC Algorith#s
Serial vs. #arallel
Stop'the'worl vs. Concurrent
Compacting vs. +on'compacting vs. Copying
23
$erial vs1 'arallel
Serial
>
2ne C#9 hanles GC tas&
#arallel
>
Multiple C#9s hanle GC tas&
simultaneously
CPU 0
CPU 1
CPU 2
CPU 3
CPU 0
CPU 1
CPU 2
CPU 3
24
$top-the-2orld vs1 Concurrent
Stop'the'worl
>
E0ecution o$ the application is completely
suspene uring GC
>
Simpler to implement (pro!
> 1onger pause time (con!
Concurrent
>
2ne or more GC tas&s can *e e0ecute
concurrently with the application
>
Shorter pause time (pro!
>
Some overhea (con!
CPU 0
CPU 1
CPU 2
CPU 3
25
Co#pacting vs1 Non-co#pacting vs1
Copying
Compacting
>
Move all live o*jects together an completely reclaim the
remaining memory
+on'compacting
>
=eleases the space .in'place/
Copying
>
Copies o*jects to a i$$erent memory area
>
The source area is empty an availa*le $or $ast an easy
su*se;uent allocations
Types of
Types of
GC Collector
GC Collector
27
Types of GC Collector
Serial Collector
#arallel Collector
#arallel Compacting Collector
Concurrent Mar&'Sweep (CMS! Collector
$erial Collector
$erial Collector
29
$erial Collector
>oth young generation GC an ol generation GC
are one serially (using a single C#9!
Stop'the'worl
9se $or most applications that
>
are running on client'style machines
>
o not have low pause time re;uirement
<e$ault $or client'style machines in ?ava SE @ A B
Can *e e0plicitly re;ueste with
>
-XX:+UseSerialGC
30
$erial Collector on &oung Generation. Before
1ive o*jects are copie $rom Een to empty survival space
=elatively young live o*jects in the occupie ()rom! survival
space are copie to empty (To! survival space while relatively
ol ones are copie to ol generation space irectly
31
$erial Collector on &oung Generation. After
>oth Een an the $ormerly occupie survival space are
empty
2nly the $ormerly empty survival space contains live o*jects
Survival spaces switch roles
32
$erial Collector on Old Generation
The ol an permanent generations are collecte
via serial mar&'sweep'compact collection algorithm
'arallel Collector
'arallel Collector
34
'arallel Collector
%lso &nown as Throughput Collector
>
9sing multiple C#9s $or young generation GC, thus increases
the throughput
Still stop'the'worl C application is suspene
9se $or ?ava applications which run on machines with a
lot o$ physical memory an multiple C#9s an o not
have low pause time re;uirement
> In$re;uent *ut potentially long ol generation GC can still occur
>
E0amples( >atch applications, *illing, payroll, scienti$ic
computing, etc
Can *e e0plicitly re;ueste with
>
-XX:+UseParallelGC
35
'arallel Collector on &oung Generation
9ses a parallel version o$ young generation collection
algorithm o$ a serial collector
36
'arallel Collector on Old Generation
The ol an permanent generations are collecte
via serial mar&'sweep'compact collection algorithm
(as in the Serial Collector!
'arallel Co#pact
'arallel Co#pact
Collector
Collector
38
'arallel Co#pact Collector
It uses a new algorithm $or ol generation collection
>
<i$$erence $rom the #arallel Collector
Still stop'the'worl
9se $or ?ava applications which run on machines with a
lot o$ physical memory an multiple C#9s an o have
low pause time re;uirement
>
#arallel operations o$ ol generation collection reuces pause
time
Can *e e0plicitly re;ueste with
>
-XX:+UseParallelOldGC
39
'arallel Co#pact Collector on &oung
Generation
It uses the same algorithm $or young generation
collection as #arallel Collector
40
'arallel Co#pact Collector on Old
Generation
Three phases
>
Mar&ing phase
>
Summary phase
>
Compaction phase
Still stop the worl
Concurrent +ar3
Concurrent +ar3
and $2eep 4C+$5
and $2eep 4C+$5
Garage Collector.
Garage Collector.
,egular +ode
,egular +ode
42
C+$ Collector
9se when an application nees shorter pause time an
can a$$or to share processor resources with GC when
application is running
>
%pplications with relatively large set o$ long'live ata (a large
ol generation! an run multi'C#9 machines
>
E0ample( We* server
%lso calle 1ow #ause Collector
Can *e e0plicitly re;ueste with
>
-XX:+UseConcMarkSweepGC
> -XX:ParallelCMSThreads=<n>
43
C+$ Collector on &oung Generation
It uses the same algorithm $or young generation
collection as #arallel Collector
44
C+$ 'hases 4for Old Generation GC5
Initial mar&
>
Stop'the'worl pause to mar& $rom roots
>
+ot a complete mar&ingDonly one level eep
Concurrent mar&
>
Mar& $rom the set o$ o*jects $oun uring Initial Mar&
=emar&
>
Stop'the'worl pause to complete mar&ing cycle
>
Ensures a consistent view o$ the worl
Concurrent Sweep
>
=eclaim ea space, aing it *ac& onto $ree lists
Concurrent =eset
45
C+$ 'hases
Java thread
GC thread
I
n
i
t
i
a
l

M
a
r
k
C
o
n
c
u
r
r
e
n
t

M
a
r
k
R
e
m
a
r
k
C
o
n
c
u
r
r
e
n
t

S
w
e
e
p
R
e

e
t
CPU 0
CPU 1
CPU 2
CPU 3
CPU !
CPU "
CPU #
CPU $
GC active
46
C+$ 2ith 'arallelis#
=emar& is typically the largest pause
>
2$ten larger than -oung GC pauses
>
#arallel remar& availa*le since ?ava @ plat$orm
Single mar&ing threa
>
Can &eep up with EFC4 cpus, usually not more
>
#arallel concurrent mar& availa*le in ?ava B plat$orm
Single sweeping threa
>
1ess o$ a *ottlenec& than mar&ing
>
#arallel concurrent sweep coming soon
47
C+$ 'hases 2ith 'arallelis#
Java thread
GC thread
I
n
i
t
i
a
l

M
a
r
k
C
o
n
c
u
r
r
e
n
t

M
a
r
k
R
e
m
a
r
k
C
o
n
c
u
r
r
e
n
t

S
w
e
e
p
R
e

e
t
CPU 0
CPU 1
CPU 2
CPU 3
CPU !
CPU "
CPU #
CPU $
GC active
48
Concurrent +ar3 $2eep Collector
Scheuling o$ collection hanle *y GC
>
>ase on statistics in ?GM
>
2r 2ccupancy level o$ tenure generation
>'HH(CMSTrigger=atio
>'HH(CMSInitiating2ccupancy)raction
Concurrent +ar3
Concurrent +ar3
and $2eep 4C+$5
and $2eep 4C+$5
Garage Collector.
Garage Collector.
Incre#ental +ode
Incre#ental +ode
50
C+$ Collector in Incre#ental +ode
CMS Collector can *e use in a moe in which the
concurrent phases are one incrementally
>
Meant to lesson the impact o$ long concurrent phases *y
perioically stopping the concurrent phase to yiel *ac&
processing to the application
Can *e e0plicitly re;ueste with
>
-XX:+CMSIncreen!alMode
51
C+$ 'hases
Java thread
GC thread
I
n
i
t
i
a
l

M
a
r
k
C
o
n
c
u
r
r
e
n
t

M
a
r
k

i
n
t
e
r
l
e
a
v
e
d

w
i
t
h

a
p
p
l
i
c
a
t
i
o
n
R
e
m
a
r
k
C
o
n
c
u
r
r
e
n
t

S
w
e
e
p
R
e

e
t
CPU 0
CPU 1
CPU 2
CPU 3
CPU !
CPU "
CPU #
CPU $
GC active
52
Other Incre#ental C+$ Options
'HH(CMSIncremental<utyCycleIJn"> (e$ault( @5!
'HH(CMSIncremental<utyCycleMinIJn"> (e$ault(
K5!
'HH(LCMSIncremental#acing (e$ault( true!
"rgono#ics
"rgono#ics
54
J/+ "rgono#ics
Throughput Collector
Ergonomics ena*les the $ollowing(
>
Throughput gar*age collector an %aptive Si3ing
>
('HH (L9se#arallelGC 'HH(L9se%aptiveSi3e#olicy!
>
Initial heap si3e o$ KMBF o$ physical memory up to KG*yte
>
Ma0imum heap si3e o$ KMF o$ physical memory up to KG*
>
Server runtime compiler ('server!
To ena*le server ergonomics on NO'*it Winows, use
the $ollowing $lags(
>
'server 'Hm0Kg 'HH(L9se#arallelGC
>
Garying the heap si3e.
55
)sing J/+ "rgono#ics
Ma0imum pause time goal
>
'HH(Ma0GC#auseMillisI<nnn>
>
This is a hint, not a guarantee
>
GC will ajust parameters to try an meet goal
>
Can aversely e$$ect application throughput
Throughput goal
>
'HH(GCTime=atioI<nnn>
>
GC Time ( %pplication time I K M (K L nnn!
>
e.g. 'HH(GCTime=atioIK6 (@7 o$ time in GC!
)ootprint goal
>
2nly consiere i$ $irst two goals are met

Gargbage Collection (GC)
Gargbage Collection (GC)
Schemes
Schemes
Sang Shin
Sang Shin
Technology Evangelist Technology Evangelist
Sun Microsystems, Inc. Sun Microsystems, Inc.
www.javapassion.com www.javapassion.com
56