Uniprocessor Garbage Collection Techniques Paul R. Wilson.

Uniprocessor Garbage Collection Techniques

Paul R. Wilson

FromSapce/Tospace before

Ft after

Too much

3mb vs. 6mb

The Two-Phase Abstraction 1. Detection

2. Reclamation

Why Garbage Collect at All? Safety

Memory leaks Continued use of freed pointers

Simplicity

Why Garbage Collect at All? Flexibility

Hard coded program limits Efficiency!

Who is responsible for deletion? Extraneous copies

Liveness and Garbage There is a root set which is defined

as live.

Anything reachable from a live pointer is also live

Everything else is garbage

The Root Set The Root Set

Static global and module variables Local Variables Variables on any activation stack(s)

Everyone else Anything Reachable From a live value

Reference Counting Advantages

Implicitly distributes garbage collection

Real Time guarantees with deferred reclamation Keep a list of zeroed objects not yet

processed Memory efficiency, can utilize all

available memory with no work room

Reference Counting Pitfalls Conservative- needs a separate GC

technique to reclaim cycles Expensive- pointer reassignment

requires: Increment Decrement Zero Check

Stack Variables frequent creation/destruction Can be optimized to some extent

Reference Counting

Ref counting, unreclaimable

Deferred Reference Counting Defer deletion of zero counted

objects Periodically scan the stack for

pointers

Mark-Sweep Collection Starting From the root set traverse

all pointers via depth/breadth first search.

Free everything that is not marked.

Non-Copying issues Same as for traditional allocators

Fragmentation Memory block size management Locality of reference- interleaved new/old

General issues- work proportional to heap size

Copying Advantages Memory locality preserved

Disadvantages Lots of copying!

“Scavenging”

Stop and Copy How to update multiple pointers to

the same object? Forwarding Pointers

Mark/Sweep is proportional to the amount of live data. Assuming this stays roughly constant, increasing memeory will increase efficiency.

Non Copying Version Facts

Allocated with a color Fragmentation

Advantages Does not require pointer rewriting

Supports obscure pointer formats, C friendly

In place collection Conservative estimates

Useful for languages like C Pointers can be safely passed to

foreign libraries not written with Garbage Collection in mind

Incremental Tracing Collectors The ‘Mutator’ The reachability graph may change

From the garbage collectors point of view the actual application is merely a coroutine ir cuncurrent process with an unfortunate tendency to modify data structures that the collector is trying to traverse

Floating Garbage Can’t survive more than one extra round

Real Time Garbage Collection Incremental Tracing Collectors

In Place Collection Many readers single writer(mutator)

As a Copying Collector Multiple Readers Multiple Writers

Tricolor Marking White

Initial color for an object subject to collection

Black Objects that will be retained after the current

round

gray Object has been reached, but not its descendents

Wave front effect

A violation of the Coloring Invariant

Read Barrier Detects an attempt to read a white

object and immediately colors it gray

Write Barrier Traps attempts to write a pointer

into an object

Some algorithms Snapshot-at-beginning write

barrier Black-only read barrier Baker’s read barrier Dijkstra’s write Barrier Steele’s write Barrier

Baker’s Read Barrier Allocates Black Grey Objects cannot be reverted to

white Immediately Invalidates fromspace Any pointer access to fromspace

causes the GC to grey the target object by copying it to tospace if necessary and updating the pointer.

Baker’s Non Copying Scheme Real Time Friendly

Treadmill

Black Only Read Barrier When a white object in fromspace

is touched it is scanned completely.

Replication Copying Collection Until copying from from space to to

space is completed, the mutator continues to read from from space.

Write updates must be trapped to update tospace.

Single simultaneous ‘flip’ where all pointers are updated.

Expensive for standard hardware, but cheap for functional languages

Real time considerations Read Barriers add an unpredictable cost

per pointer access Nilson background scavenger, reserve only

Write barrier may be more expensive overall, but the cost per access is well bounded

Guaranteeing progress allocation clock, frees per allocation

Statically allocate troublesome objects

Results Writer barrier more efficient on

standard hardware

Snapshot at the Beginning Catches pointers which try to escape from white

objects If a pointer is replace in a black object, the

replaced pointer is first stored. All overwritten pointers are saved via a write barrier. All objects that are live at the beginning of collection

remain live Allocate Black during collection round

Incremental Update Reverts black to gray when an object is written to, or

else grays they new pointed to object

Incremental Update with Write-Barrier(Dijkstra)g

Catches pointers that try to hide in black objects Reverts Black to gray If the overwritten pointer is not

pointed to elsewhere then it is garbage

Allocated white. Newly allocated objects assumed unreachable

Motivation for a new Strategy Most objects are short lived

80% to 90% die within a few million instructions

Objects that don’t die quickly are more likely to live a while

Long lived objects are copied over and over

Excessive Paging in Scanning if the heap must exceed available physical memory

Generational Garbage Collection

Generational gc before

Generational gc after

Gc memory usage

Variations of generational collection Intergenerational references

Write barrier Old to younger Young to old

Collection Advancement policies

Advance always Advance after 2 roundsCounter in the header field?Advance always? Semispace in the last generation3 spacesBucket brigadeMark compact in the oldest generation for memory efficiency