1. Rifat Shahriyar Stephen M. Blackburn Australian National University
High Performance Reference Counting and Conservative Garbage Collection
Rifat Shahriyar
Stephen M. Blackburn
Australian National University
Kathryn S. McKinley
Microsoft Research

2. Down for the Count? Getting Reference Counting Back in the Ring ISMM’12
What happened 53 years ago?

3. Why Reference Counting?
Advantages
Immediacy
Object local
Basic RC is easy
Disadvantages
Cycles
Performance

4. Can we get RC back in the ring?
Problem
One of the two fundamental GC algorithms
Many advantages
Neglected by performance-conscious VMs
So how much slower is it?
Can we get RC back in the ring?
30%

5. RC vs. MS
New RC ≈ MS

6. Summary Old RC New RC Performance 30% slower than MS
40% slower than production
New RC
Limited bit count
Optimization for new objects
Performance
Matches MS
Still 10% slower than production
< 2012
2012 6

7. Taking Off the Gloves with Reference Counting Immix OOPSLA’13

8. Why So Slow? GC
Total
Mutator

9. Looking a Little Deeper…
L1 D Cache Misses
Instructions Retired
Time
Using Managed Runtime Systems to Tolerate Holes in Wearable Memories

10. Looking a Little Deeper…
Free List
Lets see which GC uses which allocator
RC and MS – Free List
SS and Immix – Bump pointer
L1 D Cache Misses
Instructions Retired
Time
Bump Pointer
Using Managed Runtime Systems to Tolerate Holes in Wearable Memories

11. RC Immix Combines RC and Immix Exploit Immix’s opportunistic copy
Line/block reclamation
Line live object count with object reference count
Exploit Immix’s opportunistic copy
Observe new objects can be copied by first GC
Observe old objects can be copied by backup GC
Using Managed Runtime Systems to Tolerate Holes in Wearable Memories

12. 3% faster then Gen Immix, +6% worst case, -21% best case
Total time
3% faster then Gen Immix, +6% worst case, -21% best case

13. Summary RC Immix Great performance Transforms RC
-3%
RC Immix
Object-local collection
Excellent mutator locality
Copying with RC
Great performance
Outperforms fastest production
Transforms RC

14. Fast Conservative Garbage Collection OOPSLA’14
What happened 53 years ago?

15. GC is Ubiquitous GC implementations
Exact
Conservative
High performance systems use exact GC
Conservative GC is popular
roots
heap
roots
heap
heap
roots
GC – needs to find all live/dead objects
Start from the roots
Roots - all references into the heap held by runtime including stacks, registers, statics, and JNI
Conservative GC is generally used in less performant systems
exact
conservative

16. Root Conservative GC heap roots int
We are interested in root conservative GC
Where
References in the roots are not precisely known
But references in the heap objects are precisely known

17. We are interested in managed languages
Why Conservative GC
Advantages
No cooperation from compiler and runtime
Engineering accurate stack maps is challenging
Enable some compiler optimizations
Disadvantages
Must handle ambiguous references
Performance
We are interested in managed languages
Reference counting has some interesting advantages.
Our goal is to make it faster than the production.
Zoom in on the result

18. Performance of Conservative GC
BDW suffers 12% and MCC suffers 45% overhead

19. Ambiguous Reference
Pointers? – retain their referents and transitively reachable objects (Excess retention)
Values? – not modify them and pin the referents (Pinning)
Corrupt heap? – guarantee validation before updating per-object metadata (Filtering)

20. Non-moving Boehm-Demers-Weiser (BDW) widely used Problems
free-list allocator
mark-sweep trace to reclaim garbage
Problems
Free-list suffers bad locality than contiguous
With object type precision, a overly restrictive design

21. Mostly copying aka Bartlett Style with many variants
Two twists over the classic semi-space
to-space and from-space are linked lists of discontiguous pages
Promotes page referenced by ambiguous root
Problems
Semi-space suffers from huge collection cost
Space waste due to page level pinning
Objects can’t span pages and allocator can’t use pinned page

22. RC Immixcons matches production Gen Immix
Total time
RC Immixcons matches production Gen Immix

23. Summary Conservative GC New designs Conservative RC Immix
Dominated by BDW and MCC
Significant overheads
Heap org. key to performance
New designs
Low overhead object map
Immix line based pinning
Conservative RC Immix
Matches fastest production

24. Conclusion