Ulterior Reference Counting: Fast Garbage Collection without a Long Wait Author: Stephen M Blackburn Kathryn S McKinley Presenter: Jun Tao

Outline Introduction Background Ulterior Reference Counting Methodology Results Conclusion

Introduction A long-standing and unachieved goal for GC (garbage collector) – Short pause times – Excellent throughput Throughput and responsiveness of generational collector (BG-MS) and Reference counting

Introduction A new generational collector: Ulterior Reference Counting (URC) – Copying nursery – Reference counting mature space – Use a deferred reference counting – Divides the heap into logical partitions of RC and non- RC objects – Reference count only infrequently mutated fields – Defer the fields of highly mutated objects and enumerate them Trace the deferred pointers Use a write barrier to remember them

Introduction Object lifetime and mutation demographics combine well to fit that requirement – Young objects mutate frequently and die at a high rate Favor generational collection, with copying algorithm using fast contiguous allocations for the young objects – Old objects mutate infrequently and die at a slower rate Favor a space efficient free-list allocator and a reference counting algorithm

Introduction BG-RC: a hybrid generational collector – BG: bounded copying nursery – RC: Reference counts the mature space – Ignores nursery pointer mutations – For surviving nursery objects Enumerates live pointers during tracing Copies them to RC space Computes reference counts for the mature space Reclaims objects with no reference

Background Generational Collection – Use tracing to identify dead objects indirectly – Copying collector Copies all live objects into another space Works well when few objects survive Use bump-pointer allocation – Mark-sweep collector Marks live objects Frees all unmarked objects Uses free-list allocation Needs no copy reserve Poor memory utilization without compaction

Background Generational Collection – How to organize nursery collection Flexible-sized nursery Fixed-size nursery Bounded nursery

Background Reference Counting – Advantage The work of garbage detection is spread out over every mutation – Disadvantage Additional algorithms must reclaim cycles Tracking every pointer mutation is expensive

Background Reference Counting (continued) – Mechanisms Deferral – Examines certain heavily mutated pointers periodically – Deferral phase in which RCs are not correct – RC phase in which they are Buffering – Do not perform RC increments and decrements immediately – Buffer and process them in RC phase Coalescing – The periodicity of deferred RC implies that only the initial values and final values of pointer fields are relevant – Uses the differences to generate increments and decrements

Background Reference Counting (continued) – RC Formal Definitions Mutation event RCM(p) – i.e., RC(p before )--, RC(p after )++ – May be buffered or performed immediately Retain event RCR(p) Deferral – No mutation generates RCM(p) – Need a RCR(p) to preserve objects

Ulterior Reference Counting Generalizing Deferral A RC integrate event RCI(p) changes a deferred pointer to not-deferred is needed.

Ulterior Reference Counting Deferral policies – Determines for each pointer whether or not to perform mutation events – Three approaches to enumerates deferral set Trace Deferred Fields – Trace all live deferred fields just prior to every RC phase Record Mutated Deferred Fields Record Mutated Deferred Objects

Ulterior Reference Counting A Generational RC Hybrid Collector (BG-RC) – For young objects Bump-pointer allocation Copying collection – For old objects Free-list allocation Reference counting collection

Ulterior Reference Counting A Generational RC Hybrid Collector (BG-RC) (continued)

Ulterior Reference Counting A Generational RC Hybrid Collector (BG-RC) (continued) – Write Barrier Remembers pointers into the nursery from the non- nursery spaces An object remembering coalescing barrier

Ulterior Reference Counting – Write Barrier (continued)

Ulterior Reference Counting Controlling Pause Times – Pause time : Nursery collection time + reference counting time – Appropriate nursery size Large Small – Frequent collections – Diminishes the effect of coalescing in RC – Gives nursery objects less time to die – Bound the accumulation of RC work between collections Limit the growth of meta data – Modified object buffer – Decrement buffer

Ulterior Reference Counting Cycle Detection – Bacon and Rajan’s trial deletion algorithm Colors the objects from all decrements which do not go to zero purple and puts them on a list If a purple still non-zero at the end of a RC phase, computes the object and objects reachable from it gray and decrements their RCs All gray objects with RC=0 are cyclic garbage – The authors’ choice Perform cycle detection with increasing probability when the available heap space falls toward a user- defined limit

Methodology Jikes RVM and JMTk Collectors – RC: coalescing, deferred – BG-RC – MS – BG-MS: bounded copying nursery; MS mature space

Results

Sensitivity to Heap size – Generational collectors perform better on both GC time and mutator time Benefit from bump-pointers – As the heap size shrinks, BG-RC degrades more rapidly than BG-MS Pause time guidelines prevent it from reclaiming cyclic garbage promptly – Mutator time Performance: MS > BG-MS > BG-RC

Results BG-RC Sensitivity to Variations in Collection Triggers (defaults are 4MB nursery, 60ms time cap, and 512KB cycle trigger)

Conclusion Matches allocation and collection policies to the behaviors of older and younger objects – Copying collector on nursery – Reference Counting collector on mature space Achieves good performance on both throughput and responsiveness