1. Garbage-First Collector:
Current and Future
Adaptability and
Ergonomics.
Monica Beckwith Charlie Hunt John Cuthbertson

2. What's the talk about?
2
 Sneak a peek under the hood of the latest and
coolest garbage collector, Garbage-First!
 Dive deep into G1's adaptability and ergonomics
 Discuss the future of G1's adaptability

3. About the Presenters
3
 Charlie Hunt
Architect,
Performance
Engineering,
Salesforce
 Monica
Beckwith
Performance
Architect,
Servergy
 John
Cuthbertson
GC Guru,
Azul Systems

4. Agenda
 Key Concepts
 Definition
 Ergonomics
 Adaptability (if any)
 Future Adaptability Needs of G1
 Need More Information?
 Questions? - Let‟s Discuss!
4

5. Key Concepts
 Remembered Sets (RSets)
 Marking Threshold and Concurrent Cycle
 Collection Set (CSet) and Collections
 Mixed Collection
 Young Collection
 Old Regions Collection
 Humongous Allocations
 Evacuation Failures
 Reference Processing
5

6. Key Concept – Remembered Sets
6

7. Remembered Sets
 Per-region entries
 Each RSet keeps track of outside references
into its “owning” region
 RSets help regions to be independently GC‟d
 No need to scan the entire heap!
 G1 maintains RSets for –
 old-to-young and
 old-to-old references
7

8. Remembered Sets (RSets)
8
Region 2
Region 1
Region 3
RSet for
Region 1
RSet for
Region 3
RSet for
Region 2
* Figure referenced from InfoQ article: http://www.infoq.com/articles/tuning-tips-G1-GC

9. RSets – Ergonomics and Adaptability
 Concurrent Refinement threads –
 Help maintain RSets
 Concurrent updating
 Post-write Barriers –
 After a write
 Help track cross region updates
 Coarsening –
 RSets transitioning through different levels of
granularity
9

10. RSet - Concurrent Refinement
 Concurrent processing of the filled update
buffers
 Tiered deployment
 Max number can be set by
-XX:G1ConcRefinementThreads
 The processing of update buffers can eventually
fall to the mutator (application) threads
 Need to avoid such a scenario
10

11. RSet - Coarsenings
 Three levels of granularity –
 Sparse
 Hash table of card indices; Fastest to scan
 Fine
 Card indices held in a bitmap; Scan not as fast as Sparse
table
 Coarse
 One bit for each region; Slowest to scan
11

12. Key Concept – Marking Threshold
and Concurrent Cycle
12

13. Marking Threshold and Concurrent
Cycle
 Threshold default: 45% of your Java heap
 -XX:InitiatingHeapOccupancyPercent=<value>
 When threshold‟s crossed, G1 starts a
concurrent cycle
 Some phases are concurrent and some are stop-the
world
 Multi-phased concurrent marking cycle finds the
“best” regions to be collected
 Live-ness accounting
13

14. Marking Threshold and Concurrent
Cycle
 After the marking phase is complete, G1 has
information on which old regions to collect
 Regions are ordered based on “collection
efficiency”
 Expensive regions would be regions with lots of live
data and large RSets
 Completely free regions are collected during
cleanup phase
 Examples are shown later in this presentation …
14

15. Marking Threshold – Example 1
15
Default IHOP IHOP increased to 75%
Java heap size Java heap sizeMax heap occupancy
Max heap occupancy
Marking Threshold
Marking Threshold

16. Marking Threshold – Example 2
16
Default IHOP IHOP increased to
70%
25% Young GCs
60% Mixed GCs
30% Mixed GCs
64% Young GCs
Better to do more young GCs
than mixed GCs

17. Key Concept – Collection Set and
Collections
17

18. Collection Set
 A set of regions to be collected during a GC
evacuation pause
 Young Collection will have only and all young
regions in the CSet
 Mixed Collection will have both young and old
regions in the CSet
 Live data in the CSet is evacuated/copied
during a GC cycle
18

19. Collection Set – Young Collection
Look for the following in PrintAdaptiveSizePolicy enabled log –
6676.431: [GC pause (G1 Evacuation Pause) (young) 6676.431: [G1Ergonomics (CSet
Construction) start choosing CSet, _pending_cards: 147002, predicted base time: 42.39 ms,
remaining time: 157.61 ms, target pause time: 200.00 ms]
6676.431: [G1Ergonomics (CSet Construction) add young
regions to CSet, eden: 950 regions, survivors: 74 regions,
predicted young region time: 160.72 ms]
6676.431: [G1Ergonomics (CSet Construction) finish choosing CSet, eden: 950 regions,
survivors: 74 regions, old: 0 regions, predicted pause time: 203.11 ms, target pause time:
200.00 ms]
19

20. Collection Set – Mixed Collection
5884.952: [GC pause (G1 Evacuation Pause) (mixed) 5884.952: [G1Ergonomics (CSet
Construction) start choosing CSet, _pending_cards: 167183, predicted base time: 48.30 ms,
remaining time: 151.70 ms, target pause time: 200.00 ms]
5884.952: [G1Ergonomics (CSet Construction) add young regions to CSet, eden: 952 regions,
survivors: 72 regions, predicted young region time: 225.39 ms]
5884.954: [G1Ergonomics (CSet Construction) finish
adding old regions to CSet, reason: reclaimable percentage
not over threshold, old: 134 regions, max: 308 regions,
reclaimable: 1285706456 bytes (9.98 %), threshold: 10.00 %]
5884.954: [G1Ergonomics (CSet Construction) finish choosing CSet, eden: 952 regions,
survivors: 72 regions, old: 134 regions, predicted pause time: 354.55 ms, target pause time:
200.00 ms]
20

21. Collection Set – Mixed Collection
21
Young Regions
Old Regions
H Humongous Regions
✓ ✓
✓ ✓
✓ H ✓
✓
✓ H
✓
H ✓
* Figure shows both young and old collection sets.

22. Mixed Collection - Copying
22
✓ ✓
✓ ✓
✓ H ✓
✓
✓ H
✓
H ✓
Live objects are copied
into the “to-space”
regions:
- Survivor Regions
- Old Regions

23. Mixed Collection – Reclamation
23
✓ ✓
✓ ✓
✓ H ✓
✓
✓ H
✓
H ✓
G1 achieves
compaction via copying

24. CSet – Ergonomics and Adaptability
 CSet for young entirely depends on –
 The pause time target
 -XX:MaxGCPauseMillis=<value>
 All young regions are included in the CSet
 CSet for mixed collection –
 Could have a minimum number of regions based on
the MixedGCCountTarget and
 But will never cross the max number of regions set by
OldCSetThresholdPercent
24

25. CSet – Ergonomics and Adaptability
 -XX:G1MixedGCCountTarget (defaults to 8) – sets a target for
mixed collections by setting a minimum limit on the old regions
to collect per mixed collection, the goal is to not exceed the
CountTarget during the mixed collection cycle
 -XX:G1OldCSetRegionThresholdPercent (defaults to 10) – sets
an upper limit on the max number of old regions that can be
collected during a mixed collection (its expressed as a
percentage of the total heap)
25

26. CSet – Ergonomics and Adaptability
Snippets from a PrintAdaptiveSizePolicy enabled
log –
 <finish adding old regions to CSet, reason: predicted time is too high,
predicted time: 2.27 ms, remaining time: 0.00 ms, old: 12 regions,
min: 12 regions>
 <finish adding old regions to CSet, reason: old CSet region num
reached min, old: 142 regions, min: 142 regions>
 <finish adding old regions to CSet, reason: reclaimable percentage
not over threshold, old: 134 regions, max: 308 regions>
26

27. Mixed Collection – Ergonomics and
Adaptability
 Young regions – all young regions are
included in a collection
 Old regions are selected based on –
 -XX:G1MixedGCLiveThresholdPercent (defaults to 65) – sets a limit
on the live objects' occupancy such that any old region above that
threshold will not be included in any mixed collection
 Remember GC efficiency?
 -XX:G1HeapWastePercent (defaults to 10) – sets the waste target i.e.
the percentage of heap space you are willing to never collect in an
effort to avoid expensive GCs
 Remember the logs outputs that mentioned “reclaimable percentage not
over threshold”?
27

28. Young Collection – Ergonomics and
Adaptability
 Young generation size is based on your pause time
target and internally set min and max bounds
 -XX:MaxGCPauseMillis = 200 (default)
 Default min nursery size = 5% of your Java heap
 Default max nursery size = 60% of your Java heap
 Prediction logic
 Determines how much time it will take to collect 1 region
 (Re-)Sizes the young generation accordingly after each
collection
28

29. Old Regions Collection - Ergonomics
 During Mixed GCs
 Based on the criteria mentioned earlier
809.925: [G1Ergonomics (Mixed GCs) start
mixed GCs, reason: candidate old regions
available, candidate old regions: 4273 regions,
reclaimable: 111102028112 bytes (53.89 %),
threshold: 10.00 %]
29

30. Old Regions Collection - Ergonomics
 During concurrent cycle
 Entirely free (i.e. full of garbage) regions are collected
6530.615: [GC cleanup 13G->12G(18G), 0.0388540 secs]
 During Full GCs
 Collects and compacts all regions
154.725: [Full GC 1018M->369M(1024M), 1.6437640 secs]
[Eden: 0.0B(51.0M)->0.0B(51.0M) Survivors: 0.0B->0.0B Heap:
1018.3M(1024.0M)->369.4M(1024.0M)]
30

31. Key Concept – Humongous
Allocations
31

32. Humongous Objects –What Are
They?
 Objects that span >= 50% of G1‟s region size
 Ideally –
 Not that many in number
 Are long lived
 Allocated directly into the old generation into
Humongous Regions
 Avoids unnecessary copying back and forth and
expensive promotions
 Larger objects will need contiguous regions
32

33. Humongous Objects – WWG1D?
33
G1 Region –
Young Generation
G1 Region –
Old Generation
Object 1 < 50% of
G1 Region
Object 2 == 50%
of G1 Region
Object 3 > 50%
of G1 Region
Object 4 >
G1 Region
?

34. Humongous Objects – WWG1D?
34
G1 Region –
Young Generation
G1 Region –
Old Generation
Object 1 < 50% of
G1 Region
Object 2 == 50%
of G1 Region
Object 3 > 50%
of G1 Region
Object 4 >
G1 Region
* Object 4 will need contiguous regions

35. Object 4 - Humongous
Humongous Objects
35
Object 1 – NOT Humongous
Object 2 - Humongous
Object 3 - Humongous
* Object 4 needs contiguous regions
Wasted
Space

36. Humongous Objects – What’s The
Catch?
 As of 7u40, initial heap size (-Xms) determines
the region size
 G1 strives for 2048 regions
 Region size a factor of 2 ranging from 1M to 32M
 If there is a vast difference between the initial
and the max heap, normal objects may seem
humongous to G1!
36

37. Humongous Objects – What’s The
Catch?
 Let‟s look at this PrintAdaptiveSizePolicy enabled
log snippet (Note: G1 region size was 4MB):
 1361.680: [G1Ergonomics (Concurrent Cycles)
request concurrent cycle initiation, reason:
occupancy higher than threshold, occupancy:
1459617792 bytes, allocation request: 4194320
bytes, threshold: 1449551430 bytes (45.00 %),
source: concurrent humongous allocation]
37

38. Humongous Objects – What’s The
Catch?
 Notes from earlier snippet –
 Concurrent cycle requested
 Reason: Occupancy was higher than threshold
 Allocation size was 4194320 bytes
 Greater than 4MB, hence humongous allocation
 Notes from the log (not shown here) –
 Too many humongous allocations
 Concurrent cycles can‟t keep up with the allocations
 Resulting in to-space exhausted messages and eventually Full GCs
38

39. Humongous Objects – How to “Fix”
it?
 Let‟s find a region size that can accommodate
the humongous objects as regular allocations
 So the next region size up would be 8MB
 But, 4.000015MB > 50% of 8MB
 So, go to the next size up. i.e. 16MB
 Solution – Set your region size to 16MB
 -XX:G1HeapRegionSize=16M
39

40. Humongous Allocation –
Ergonomics
 Humongous Regions are not included in a
mixed collection
 Dead Humongous objects are collected during
cleanup and during Full GC
6569.877: [GC cleanup 6708M- >6384M(12G),
0.0181200 secs]
 Live Humongous objects are compacted through
during full GC
40

41. Key Concept – Evacuation Failures
41

42. Evacuation Failures
 Evacuation failures indicate that G1 ran out of
heap regions either –
 while copying to survivor regions or
 while promoting or copying live objects in-to the
old generation
 Prior to Java 7u40 evacuation failures shown as
a “to-space overflow” in the GC logs
 Java 7u40 onwards shows “to-space exhausted”
in the GC logs
42

43. Evacuation Failures – How to Avoid
Them?
 Get a baseline with bare minimum options:
 -Xmx, -Xms and -XX:MaxGCPauseMillis=<value>
 Over-tuning is NOT for G1
 Look at the output of PrintAdaptiveSizePolicy
 Too many humongous allocations?
 Increase G1HeapRegionSize
43

44. Evacuation Failures
 Plot the heap utilization stats from the log
 Marking threshold too high?
 Can‟t keep up with promotions
 Marking threshold too low?
 Not reclaiming much space from marking cycle
 Concurrent cycles taking a long time to
complete?
 Increase the thread count: ConcGCThreads
44

45. Evacuation Failures
 Sometimes survivor space gets exhausted
 Increase the G1ReservePercent
 It‟s a false ceiling
 Defaults to 10
 G1 will cap it off at 50%
45

46. Key Concept – Reference Processing
46

47. Reference Processing
47
0
10
20
30
40
50
60
70
80
90
100
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
User Time (in secs) System Time (in secs)
Real Time (in secs) Scaling

48. Reference Processing
48
9185.651: [GC remark 9185.653: [GC ref-proc, 5.2216490 secs], 5.2946850 secs]
[Times: user=5.62 sys=0.00, real=5.29 secs]
[Other: 339.7 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 287.8 ms]
[Ref Enq: 5.4 ms]
[Free CSet: 6.9 ms]
This is not good!

49. Parallel Reference Processing
49
0
10
20
30
40
50
60
70
80
90
100
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
[GCpause
remark
[GCpause
User Time (in secs) System Time (in secs)
Real Time (in secs) Scaling
After enabling -XX:+ParallelRefProcEnabled

50. Parallel Reference Processing
50
9095.305: [GC remark 9095.309: [GC ref-proc, 0.4888450
secs], 0.5621620 secs]
[Times: user=7.84 sys=0.03, real=0.56 secs]
[Other: 91.0 ms]
[Choose CSet: 0.0 ms]
[Ref Proc: 43.9 ms]
[Ref Enq: 1.0 ms]
[Free CSet: 7.4 ms]
Ooh yeah,
much better!

51. Future Adaptability
51

52. Future Adaptability
 Adaptive marking threshold?
 A static value doesn‟t cut it
 CMS has adaptive marking threshold with static
override and static max value
 Wouldn‟t it be nice if G1 had an adaptive threshold
and adaptive max value?
 Adaptive region size?
 Will help alleviate the issues that applications with
numerous short-lived “humongous” objects encounter
52

53. Future Adaptability
 Avoiding evacuation failures
 http://bugs.sun.com/bugdatabase/view_bug.do?bug
_id=8014019
 Reducing the cost of evacuation failures
 http://bugs.sun.com/bugdatabase/view_bug.do?bug
_id=8003237
 http://bugs.sun.com/bugdatabase/view_bug.do?bug
_id=8003235
53

54. Future Adaptability
 Improving RSets
 http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7187490
 Improving concurrent refinement thresholds to
reduce the possibility of long RSet updating times.
 Smarter/ adaptable criteria for selecting CSet
for old regions
 Smarter/ adaptable Mixed Collections
 http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7173711
54

55. Want More Info?
55

56. G1 Articles on the Web
 http://www.infoq.com/articles/G1-One-Garbage-Collector-To-
Rule-Them-All
 http://www.infoq.com/articles/tuning-tips-G1-GC
 http://www.oracle.com/technetwork/articles/java/g1gc-
1984535.html
 http://www.infoq.com/presentations/java-g1
 https://oracleus.activeevents.com/connect/sessionDetail.ww?SE
SSION_ID=6583
 https://blogs.oracle.com/javatraining/entry/getting_started_with
_the_g1
56

57. Questions?
Let‟s discuss „em!
57

58. More Questions?
HOL5429: Tuning Low-Pause Garbage Collection – Tue 10 am
CON3754: G1 GC: Migration to, Expectations, and Advanced Tuning –
Wed 10 am
CON7624: Understanding Java Garbage Collection and What You Can
Do About It. – Wed 11:30 am
GC Tuning BOF4020 - Tonight @7:30 pm
JVM Performance BOF4471 – Tonight @8:30 pm
Email: hotspot-gc-use@openjdk.java.net & hotspot-gc-
dev@openjdk.java.net
58