September 11, 2003

Beltway:Getting Around GC Gridlock

Steve Blackburn, Kathryn McKinley

Richard Jones, Eliot Moss

September 11, 2003

Where is GC?• Renewed interest in GC– Object-oriented languages with automatic memory

management (Java, C#) • Major outstanding issues– GC throughput--10 to 70% penalty• Beltway - high throughput with new copying

framework– Combing high throughput and low pause times• Ulterior Reference Counting

– Understanding & exploiting locality with GC– Dynamically adaptive GC

September 11, 2003

Where is GC?

September 11, 2003

Beltway

• A new framework for GC algorithms– Subsumes existing copying GCs– Includes new, faster copying GCs

• Novelty– Generality – New algorithms– New GC mechanisms

September 11, 2003

Object Demographic Observations

• Lifetimes– Young objects: Most very short lived• Infant mortality: ~90% die young (within 4MB of

alloc)– Old objects: most very long lived (bimodal)• Mature morality: ~5% die each 4MB of new alloc

• Pointer mutations– Older to younger pointers across many objects are rare• less than 1%

– Most mutations among young objects• 92 to 98% of pointer mutations

September 11, 2003

5 Key Ideas in Copying GC

• Generational hypothesis (Ungar, Lieberman/Hewitt)–Young objects: most die, so collect frequently–Old objects: most live, so collect infrequently

• Older-first principle (Stefanovic/McKinley/Moss)–Give objects as much time to die as possible

• Incrementality improves responsiveness• Copying GC can improve locality

Can one GC incorporate them all?

September 11, 2003

Two Organizational Principles

• Increments– Independently collectable regions

• Belts– FIFO groupings of increments

September 11, 2003

• Notice:– Need for copy reserve (rightmost increment)– Incompleteness (w.r.t. cross-increment cycles)– Corresponds to “OF-Mix” GC [Stefanovic 99]– Semi-Space (degenerate form, 2 increments)

A Simple Example

0 1 2 3 4 5 6 7

Increments

Belt

0 1 2 3 4 5 6 71 2 3 4 5 6 7 82 3 4 5 6 7 8 9

September 11, 2003

? Rules:–Only N increments available (N=8 here)–Must reserve one for copying– Always collect leftmost, lowest full

increment

A More Interesting Example

• Notice:– Generational & older first principles– Still incomplete

• This heap organization is Beltway X.X– X is the increment size (e.g. Beltway 20.20 or Beltway

14.14)• Beltway 100.100 = Appel-style generational

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

1 2 3 4 5 6 7 8

etc. etc. …0 1 2 3 4 5 6 7

33 34 35 36 37 38 39 40

1 2 3 4 5 6 7 82 3 4 5 6 7 8 9

34 35 36 37 38 39 40 41

September 11, 2003

Completeness

• Add a third belt (“Beltway X.X.100”)– Large increment (100% of available)– Only collect third belt when it is full• Reduce incrementality, gain completeness

3 4 5 6 7 8 9 10

33 34 35 36 37 38 39 40

0 1

September 11, 2003

Efficient Implementation

• Crucial issues:– Write barrier cost (consequence of incrementality)– Knowing when to collect– Minimizing ‘copy reserve’ overhead

September 11, 2003

New Mechanisms

• Frames– Incrementality with cheap w/b– One for each <source, target> frame pair• 2n-aligned contiguous virtual memory space• Accommodates an increment*• Allows fast inter-increment w/b

7x2f 8x2f 9x2f 10x2f

frameincrement

2f

• Dynamic, conservative copy reserve– Only reserve as much as necessary

• GC triggers– Don’t necessarily collect only when full

• Remembered sets– One for each <source, target> frame pair

September 11, 2003

Results

• Implemented in Jikes RVM with GCTk• Compare to state of the art copying GC

- Appel-style generational• Measure over 33 heap sizes– 1.0 through to 3.0 x minimum heap size– GC time– Total execution time

• Normalize results to best

September 11, 2003

GC Time(geometric mean of 6 benchmarks)

September 11, 2003

Execution Time(geometric mean of 6 benchmarks)

September 11, 2003

Execution Timepseudojbb

September 11, 2003

Beltway: Conclusions

• Beltway: a new framework– New family of GC algorithms– Subsumes existing copying GCs– Opens up new GC possibilities

• Generality + efficiency– Better, faster GCs

September 11, 2003

Remaining Problem: Combining Throughput &

Responsiveness_228_jack

00.10.20.30.40.50.60.70.80.9

1

1.00E+01 1.00E+02 1.00E+03 1.00E+04Window (msec)

Utilization

FG-MSRC

September 11, 2003

Discussion Questions

• Worst case space overhead?• Pause time in a 100 belt increment?• Fixed-size increments?• What happens if the remsets get big?• Locality?