A Comparative Evaluation of Parallel Garbage Collectors

1
A Comparative Evaluation ofParallel Garbage
Collectors

• Dick Attanasio David F. Bacon
• Anthony Cocchi Stephen Smith
• IBM T.J. Watson Research Center
• Presented at the University of Washington
• June 5, 2001

2
Jalapeño Goals

• Java VM for large MP servers
• Scalability
• Large heaps
• Large numbers of processors
• Large numbers of threads
• Example IBM RS/6000 S80 running Websphere
• 50 GB RAM
• 24 64-bit PowerPC RS64 III Processors
• Thousands of servlets/beans
• GC must be parallel and/or concurrent

3
The Problem with Parallel GC

• Limited experience
• Logic programming, functional languages
• Applicability to Java?
• Unknown relative performance of techniques
• Scaling properties not known
• Processor scaling
• some studies
• Memory scaling
• zilch

4
Compare Parallel GCs

• Implement a wide variety of techniques
• Gain first-hand experience
• Both general and Java-specific issues
• Short term
• Find best GC and use as default
• Long term
• Understand how to target GCs to apps/desiderata
• Select automatically or even adaptively

5
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

6
Jalapeño Features

• Entirely written in Java
• M x N threading
• Java threads multiplexed onto virtual processors
• Processor-local allocation
• Small requests handled from VP-local blocks
• Safe points
• Context switch at compiler-controlled points
• Type accuracy
• Exact GC
• Inlined allocation
• Just another inlined Java method

7
Modular Collector Design

• Multiple compilers (base, quick, opt)
• Multiple GCs
• Shared code
• Parallel collection phases and synchronization
• Stack/object scanning
• Large object management
• GC must be selected at build time
• Bootstrapping issues

8
Garbage Collectors Implemented

• Copying semi-space (CP)
• Non-compacting mark-sweep (MS)
• Generational
• Copying (CG)
• Mark-sweep (MSG)
• Hybrid copying young, mark-sweep old (H)
• Concurrent reference counting/cycle collecting
• Described in PLDI01 and ECOOP01 papers

9
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

10
Memory Layout
11
Large Object Space
Small Object Space
Boot Image

• Uses first-fit strategy with 4KB pages
• For objects larger than 2KB
• Objects never move
• Large object code shared by all collectors

12
Semi-space Heap Layout
SEMISPACE 1
SEMISPACE 2
Boot Image
Large Object Space
Allocated Objects
13
Generational Semi-space Layout
SEMISPACE 1
SEMISPACE 2
NURSERY
Boot Image
Large Object Space
Allocated Objects
14
Mark-sweep Heap Layout
Boot Image
Large Object Space
15
Mark-sweep Heap Layout
Boot Image
Large Object Space
Mark Arrays
16
Hybrid Heap Layout
NURSERY
MATURE SPACE
Boot Image
Large Object Space
17
Allocation

• Inlined automatically
• all Java code
• Properties analyzed at compile-time
• Object size
• Finalizable
• Allocation cost
• 10 PowerPC instructions for copying
• 17 PowerPC instructions for mark-sweep

18
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

19
General Properties of Collectors

• Parallel
• Stop-the-world
• Phases separated by barrier synchronization
• Work buffers for scanning
• Load balancing
• Since its in Java, copying collector
• Must move stack, register save objects

20
Alloc Fails
A
B
C
D
E
F
VP1
VP2
VP3
Init.
Find Roots
Mark/Copy All Reachable Objects
Find Final.
Scan Final.
End GC
Mutator Activity
Garbage Collection
21
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

22
Performance Factors Winner(s)

• Object allocation CP H CG
• Copying MS H CG
• Frequency of Collection MS H
• Space loss
• Semispace (mature, nursery) MS H
• Fragmentation CP
• Write barriers CP MS
• Locality CP

23
Performance Fastest
24
Performance Smallest
25
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

26
SPECjbb 400 MB Processor Scaling
27
SPECjbb 8 CPUs Heap Scaling
28
SPECjbb Total Mature GC Time
29
SPECjbb Total GC Time
30
SPECjbb Mutator-to-GC Switch
31
SPECjbb Major Collection Time
32
SPECjbb Minor Collection Time
33
Outline

• Motivation
• Overview of Collectors
• Memory Organization
• Phases of Collection
• Performance overview 6 benchmarks
• Performance details SPECjbb
• Conclusions

34
Conclusions

• JVM in Java wins
• M x N threading allows fast GC transition
• Exact GC
• When memory is abundant
• Winner is not obvious
• When memory is constrained
• Mark-sweep or hybrid always wins

35
How to choose GC dynamically?

• Memory headroom
• Working set size
• Allocation rate

36
Current Status

• Mature PowerPC compiler
• Optimizer
• Feedback-directed optimization
• Six garbage collectors
• Linux/Intel port underway
• Baseline compiler due 6/01 opt 12/01
• Other unofficial ports
• Win32
• Linux/Macintosh

37
Availability

• University license available. Licensees
• Michigan State U. Kent
• Purdue U. U. Massachussets
• Rutgers U. U. New Mexico
• U. Colorado U. Wisconsin
• U. Illinois
• More information
• dfb_at_watson.ibm.com
• http//www.research.ibm.com/jalapeno