Concurrent Cycle Collection in Reference Counted Systems

Abstract

Automatic storage reclamation via reference counting has important advantages, but has always suffered from a major weakness due to its inability to reclaim cyclic data structures.
We describe a novel cycle collection algorithm that is both concurrent — it is capable of collecting garbage even in the presence of simultaneous mutation — and localized—it never needs to perform a global search of the entire data space. We describe our algorithm in detail and present a proof of correctness. We have implemented our algorithm in the Jalapeño Java virtual machine as part of the Recycler, a concurrent multiprocessor reference counting garbage collector that achieves maximum mutator pause times of only 6 milliseconds. We present measurements of the behavior of the cycle collection algorithm over a set of eight benchmarks that demonstrate the effectiveness of the algorithm at finding garbage cycles, handling concurrent mutation, and eliminating global tracing.

Full text

Proc. European Conf. on Object-Oriented Programming, June, 2001, LNCS vol. 2072
Concurrent Cycle Collection
in Reference Counted Systems
David F. Bacon and V.T. Rajan
IBM T.J. Watson Research Center
P.O. Box 704, Yorktown Heights, NY 10598, U.S.A.
dfb@watson.ibm.com vtrajan@us.ibm.com
Abstract. Automatic storage reclamation via reference counting has important
advantages, but has always suffered from a major weakness due to its inability to
reclaim cyclic data structures.
We describe a novel cycle collection algorithm that is both concurrent — it is
capable of collecting garbage even in the presence of simultaneous mutation —
and localized — it never needs to perform a global search of the entire data space.
We describe our algorithm in detail and present a proof of correctness.
We have implemented our algorithm in the Jalape˜no Java virtual machine as part
of the Recycler, a concurrent multiprocessor reference counting garbage collector
that achieves maximum mutator pause times of only 6 milliseconds. We present
measurements of the behavior of the cycle collection algorithm over a set of eight
benchmarks that demonstrate theeffectiveness of thealgorithm at finding garbage
cycles, handling concurrent mutation, and eliminating global tracing.
1 Introduction
Forty years ago, two methods of automatic storage reclamation were introduced: ref-
erence counting [7] and tracing [23]. Since that time tracing collectors and their vari-
ants (mark-and-sweep, semispace copying, mark-and-compact) have been much more
widely used due to perceived deficiencies in reference counting.
Changes in the relative costs of memory and processing power, and the adoption
of garbage collected languages in mainstream programming (particularly Java) have
changed the landscape. We believe it is time to take a fresh look at reference counting,
particularly as processor clock speeds increase while RAM becomes plentiful but not
significantly faster. In this environment the locality properties of reference counting
are appealing, while the purported extra processing power required is likely to be less
relevant.
At the same time, Java’s incorporation of garbage collection has thrust the problem
into the mainstream, and large, mission critical systems are being built in Java, stressing
the flexibility and scalability of the underlying garbage collection implementations. As
a result, the supposed advantages of tracing collectors — simplicity and low overhead
— are being eroded as they are being made ever more complex in an attempt to address
the real-world requirements of large and varied programs.
Furthermore, the fundamental assumption behind tracing collectors, namely that it
is acceptable to periodically trace all of the live objects in the heap, will not necessarily
scale to the very large main memories that are becoming increasingly common.

2 David F. Bacon and V.T. Rajan
There are three primary problems with reference counting, namely:
1. storage overhead associated with keeping a count for each object;
2. run-time overhead of incrementing and decrementing the reference count each time
a pointer is copied; and
3. inability to detect cycles and consequent necessity of including a second garbage
collection technique to deal with cyclic garbage.
The inability to collect cycles is generally considered to be the greatest weakness
of reference counting collectors. It either places the burden on the programmer to break
cycles explicitly, or requires special programmingidioms, or requires a tracing collector
to collect the cycles.
In this paper, we present first a synchronous and then a concurrent algorithm for the
collection of cyclic garbage in a reference counted system. The concurrent algorithm is
a variantof the synchronousalgorithm with additionaltests tomaintain safety properties
that could be undermined by concurrent mutation of the data structures.
Like algorithms based on tracing (mark-and-sweep, semispace copying, and mark-
and-compact) our algorithms are linear in the size of the graph traced. However, our
algorithms are able to perform this tracing locally rather than globally, and often trace
a smaller subgraph.
These algorithms have been implemented in a new reference counting collector,
the Recycler, which is part of the Jalape˜no Java VM [1] implemented at the IBM T.J.
Watson Research Center. Jalape˜no is itself written in Java.
In concurrentlypublished work[3] we describethe Recycler asa whole,and provide
measurements showing that our concurrent reference counting system achieves maxi-
mum measured mutator pause times of only 6 milliseconds. End-to-end execution times
are usually comparable to those of a parallel (but non-concurrent)mark-and-sweep col-
lector, although there is occasionally significant variation (in both directions).
In this paper we concentrate on describing the cycle collection algorithm in suffi-
cient detail that it can be implemented by others, and give a proof of correctness which
gives further insight into how and why the concurrentalgorithm works. We also provide
measurements of the performance of the cycle collection algorithms for a suite of eight
Java benchmarks.
The rest of the paper is organized as follows: Section 2 describes previous ap-
proaches to cycle collection; Section 3 describes our synchronous algorithm for col-
lection of cyclic garbage; Section 4 then presents our concurrent cycle collection al-
gorithm. Section 5 contains proofs of correctness for the concurrent cycle collection
algorithms. Section 6 presents our measurements of the effectiveness of the algorithms.
Section 7 describes related work on concurrent garbage collection. Finally, we present
our conclusions.
Subsections 3.1 and 4.4 contain detailed pseudocode of the algorithms and can be
skipped on a first reading of the paper.
2 Previous Work on Cycle Collection
Previous work on solving the cycle collection problem in reference counted collectors
has fallen into three categories:

Concurrent Cycle Collection in Reference Counted Systems 3
– special programming idioms, like Bobrow’s groups [5], or certain functional pro-
gramming styles;
– use of an infrequently invoked tracing collector to collect cyclic garbage [8]; or
– searching for garbage cycles by removing internal reference counts [6, 22].
An excellent summary of the techniques and algorithms is in chapter 3 (“Reference
Counting”) of the book by Jones and Lins [17]. The first algorithm for cycle collection
in a reference counted system was devised by Christopher [6]. Our synchronous cycle
collection algorithm is based on the work of Mart´ınez et al [22] as extended by Lins
[20], which is very clearly explained in the chapter of the book just mentioned.
There are two observationsthat are fundamentalto these algorithms. The first obser-
vation is that garbagecycles can only be created when a reference count is decremented
to a non-zero value— ifthe referencecountis incremented,no garbageis beingcreated,
and if it is decremented to zero, the garbage has already been found. Furthermore,since
reference counts of one tend to predominate, decrements to zero should be common.
The second observation is that in a garbage cycle, all the reference counts are in-
ternal; therefore, if those internal counts can be subtracted, the garbage cycle will be
discovered.
As a result, when a reference count is decremented and does not reach zero, it is
considered as a candidate root of a garbage cycle, and a local search is performed. This
is a depth-first search which subtracts out counts due to internal pointers. If the result is
a collection of objects with zero reference counts, then a garbage cycle has been found
and is collected; if not, then another depth-first-search is performed and the counts are
restored.
Lins [20] extended the original algorithm to perform the search lazily by buffering
candidate roots instead of exploringthem immediately.This has two advantages. Firstly,
after a time, the reference count of a candidate root may reach zero due to other edge
deletions, in which case the node can simply be collected, or the reference count may
be re-incremented due to edge additions, in which case it may be ignored as a candidate
root. Secondly, it will often prevent re-traversal of the same node.
Unfortunately,in the worst case Lins’ algorithm is quadratic in the size of the graph,
as for example in the cycle shown in Figure 1. His algorithm considers the roots one at
a time, performing the reference count subtraction and restoration passes for that root
before moving on.
Therefore, Lins’ algorithm will perform a complete scan from each of the candidate
roots until it arrives at the final root, at which point the entire compound cycle will be
collected.
3 Synchronous Cycle Collection
In this section we describe our synchronous cycle collection algorithm, which ap-
plies the same principles as those of Mart´ınez et al and Lins, but which only requires
worst-case time for collection (where is the number of nodes and is the
number of edges in the object graph), and is therefore competitive with tracing garbage
collectors.

4 David F. Bacon and V.T. Rajan
2
2
2
1
5RRW%XIIHU
Fig.1. Example of compound cycle that causes Lins’ algorithm to exhibit quadratic complexity.
We also improve the practicality of the algorithm by allowing resizing of collected
objects, and show how significant constant-time improvements can be achieved by rul-
ing out inherently acyclic data structures.
Our synchronous algorithm is similar to Lins’ algorithm: when reference counts
are decremented, we place potential roots of cyclic garbage into a buffer called Roots.
Periodically, we process this buffer and look forcycles by subtracting internal reference
counts.
There are two major changes that make the algorithm linear time: first of all, we add
a buffered flag to every object, which is used to prevent the same object being added to
the root buffer more than once per cycle collection. This in turn places a linear bound
on the size of the buffer.
Secondly,we analyze the entiretransitiveclosure of Roots as a single graph,rather
than as a set of graphs. This means that the complexity of the algorithm is limited by
the size of that transitive closure, which in turn is limited by
(since Roots is
bounded by by the use of the buffered flag). Of course, in practice we hope that the
transitive closure will be significantly smaller.
In practice we found that the first change (the use of the buffered flag) made al-
most no difference in the running time of the algorithm; however, the second change
(analyzing the entire graph at once) made an enormous difference in run-time. When
we applied Lins’ algorithm unmodified to large programs, garbage collection delays
extended into minutes.
3.1 Pseudocode and Explanation
We now present detailed pseudocode and an explanation of the operation of each pro-
cedure in the synchronous cycle collection algorithm.
In addition to the buffered flag, each object contains a color and a reference count.
For an object T these fields are denoted buffered(T), color(T), and RC(T). In
the implementation, these quantities together occupy a single word in each object.

Concurrent Cycle Collection in Reference Counted Systems 5
Color Meaning
Black In use or free
Gray Possible member of cycle
White Member of garbage cycle
Purple Possible root of cycle
Green Acyclic
Red Candidate cycle undergoing
-computation
Orange Candidate cycle awaiting epoch boundary
Table 1. Object Colorings for Cycle Collection. Orange and red are only used by the concurrent
cycle collector and are described in Section 4.
All objects start out black. A summary of the colors used by the collector is shown
in Table 1. The use of green (acyclic) objects will be discussed below.
The algorithm is shown in Figure 2. The procedures are explained in detail below.
Increment and Decrement are invoked externally as pointers are added, removed,
or overwritten. CollectCycles is invoked either when the root buffer overflows,
storage is exhausted, or when the collector decides for some other reason to free cyclic
garbage. The rest of the procedures are internal to the cycle collector. Note that the
procedures MarkGray, Scan, and ScanBlack are the same as for Lins’ algorithm.
Increment(S) When a reference to a node S is created, the reference count of T
is incremented and it is colored black, since any object whose reference count was
just incremented can not be garbage.
Decrement(S) When a reference to a node S is deleted, the reference count is
decremented. If the reference count reaches zero, the procedure Release is in-
voked to free the garbage node. If the referencecount does not reach zero, the node
is considered as a possible root of a cycle.
Release(S) When the reference count of a node reaches zero, the contained point-
ers are deleted, the object is colored black, and unless it has been buffered, it is
freed. If it has been buffered, it is in the Roots buffer and will be freed later (in
the procedure MarkRoots).
PossibleRoot(S) When the reference count of S is decremented but does not
reach zero, it is considered as a possible root of a garbage cycle. If its color is
already purple, then it is already a candidate root; if not, its color is set to purple.
Then thebuffered flag is checkedto see if it has been purple since we last performed
a cycle collection. If it is not buffered, it is added to the buffer of possible roots.
CollectCycles() When the root buffer is full, or when some other condition,
such as low memory occurs, the actual cycle collection operation is invoked. This
operation has threephases: MarkRoots, which removes internal reference counts;
ScanRoots, which restores reference counts when they are non-zero; and finally
CollectRoots, which actually collects the cyclic garbage.
MarkRoots() The marking phase looks at all the nodes S whose pointers have
been stored in the Roots buffer since the last cycle collection. If the color of
the node is purple (indicating a possible root of a garbage cycle) and the reference

6 David F. Bacon and V.T. Rajan
Increment(S) ScanRoots()
RC(S) = RC(S) + 1 for S in Roots
color(S) = black Scan(S)
Decrement(S) CollectRoots()
RC(S) = RC(S) - 1 for S in Roots
if (RC(S) == 0) remove S from Roots
Release(S) buffered(S) = false
else CollectWhite(S)
PossibleRoot(S)
MarkGray(S)
Release(S) if (color(S) != gray)
for T in children(S) color(S) = gray
Decrement(T) for T in children(S)
color(S) = black RC(T) = RC(T) - 1
if (! buffered(S)) MarkGray(T)
Free(S)
Scan(S)
PossibleRoot(S) if (color(S) == gray)
if (color(S) != purple) if (RC(S) > 0)
color(S) = purple ScanBlack(S)
if (! buffered(S)) else
buffered(S) = true color(S) = white
append S to Roots for T in children(S)
Scan(T)
CollectCycles()
MarkRoots() ScanBlack(S)
ScanRoots() color(S) = black
CollectRoots() for T in children(S)
RC(T) = RC(T) + 1
MarkRoots() if (color(T) != black)
for S in Roots ScanBlack(T)
if (color(S) == purple)
and RC(S) > 0 CollectWhite(S)
MarkGray(S) if (color(S) == white
else and ! buffered(S))
buffered(S) = false color(S) = black
remove S from Roots for T in children(S)
if (RC(S) == 0) CollectWhite(T)
Free(S) Free(S)
ScanRoots()
for S in Roots
Scan(S)
Fig.2. Synchronous Cycle Collection

Concurrent Cycle Collection in Reference Counted Systems 7
count has not become zero, then MarkGray(S) is invokedto performa depth-first
searchin which the reached nodes are colored gray and internal referencecounts are
subtracted. Otherwise, the node is removedfrom the Roots buffer,the buffered
flag is cleared, and if the reference count is zero the object is freed.
ScanRoots() For each node S that was considered by MarkGray(S), this proce-
dure invokes Scan(S) to either color the garbage subgraph white or re-color the
live subgraph black.
CollectRoots() After the ScanRoots phase of the CollectCycles proce-
dure, any remaining white nodes will be cyclic garbage and will be reachable from
the Roots buffer. This procedure invokes CollectWhite for each node in the
Roots buffer to collect the garbage; all nodes in the root buffer are removed and
their buffered flag is cleared.
MarkGray(S) This procedure performs a simple depth-first traversal of the graph
beginning at S, marking visited nodes gray and removing internal reference counts
as it goes.
Scan(S) If this procedure finds a gray object whose reference count is greater than
one, then that object and everything reachable from it are live data; it will therefore
call ScanBlack(S) in order to re-color the reachable subgraph and restore the
reference counts subtracted by MarkGray. However, if the color of an object is
gray and its reference count is zero, then it is colored white, and Scan is invoked
upon its children. Note that an object may be colored white and then re-colored
black if it is reachable from some subsequently discovered live node.
ScanBlack(S) This procedure performs the inverse operation of MarkGray, vis-
iting the nodes, changing the color of objects back to black, and restoring their
reference counts.
CollectWhite(S) This procedure recursively frees all white objects, re-coloring
them black as it goes. If a white object is buffered, it is not freed; it will be freed
later when it is found in the Roots buffer.
3.2 Acyclic Data Types
A significant constant-factor improvement can be obtained for cycle collection by ob-
serving that some objects are inherently acyclic. We speculate that they will comprise
the majority of objects in many applications. Therefore, if we can avoid cycle collec-
tion for inherently acyclic objects, we will significantly reduce the overhead of cycle
collection as a whole.
In Java, dynamic class loading complicates the determination of inherently acyclic
data structures. We have implemented a very simple scheme as part of the class loader.
Acyclic classes may contain:
– scalars;
– references to classes that are both acyclic and final; and
– arrays of either of the above.
Our implementation marks objects whose class is acyclic with the special color
green. Green objects are ignored by the cycle collection algorithm, except that when

8 David F. Bacon and V.T. Rajan
Black
Purple
White
Gray
Green
DF\FOLFQHZ
GHF
:

LQF
GHF!
PDUN
GHF!
GHF
>

LQF
LQFGHF:

VFDQ
PDUN
Orange
FROOHFW
GHF
:

Red
LQFIUHH

Fig.3. State transition graph for cycle collection.
a dead cycle refers to green objects, they are collected along with the dead cycle. For
simplicity of the presentation, we have not included consideration of green objects in
the algorithms in this paper; the modifications are straightforward.
While our determination of acyclic classes is very simple, it is also very effective,
usually reducing the objects considered as roots of cycles by an order of magnitude, as
will be shown in Section 6. In a static compiler, a more sophisticated program analysis
could be applied to increase the percentage of green objects.
4 Concurrent Cycle Collection
We now describe a concurrent cycle collection algorithm based on the principles of our
synchronous algorithm of the previous section.
For the purposes of understanding the cycle collection algorithm, the multiproces-
sor reference counting system can be viewed very abstractly as follows: as mutators
create and destroy references to objects (either on the stack or in the heap), correspond-
ing increment and decrement operations are enqueued into a local buffer, called the
mutation buffer. Periodically, the mutators send these mutation buffers to the collector,
which applies the reference count updates, frees objects whose counts drop to zero, and
periodically also performs cycle collection.
Time is divided into epochs, and each mutator must transfer its mutation buffer to
the collector exactly once per epoch. However, aside from this requirement, in normal
operation no synchronization is required between mutators and the collector.
When all mutation buffers for an epoch have been transferred to the collector, the
increments for the just-completed epoch are applied; however, the decrements are not
applied until the next epoch boundary. This prevents freeing of live data which might

Concurrent Cycle Collection in Reference Counted Systems 9
otherwise occur due to race conditions between the mutators. The advantage of this
approach is that it is never necessary to halt all the mutators simultaneously.
In its implementation, the Recycler only tracks pointer updates to the heap, and
snapshots pointersin the stack at epochboundaries.Ouralgorithmis similarto Deutsch-
Bobrow deferred reference counting [9], but is superior in a number of important re-
spects. Our implementation of concurrent reference counting is most similar to the
reference counting collector of DeTreville [8]. The Recycler is described in detail by
Bacon et al [3].
4.1 Two Phase Cycle Collection
Now that we have abstracted the concurrent system to a collection of mutators emit-
ting streams of increment and decrement operations, and a reference counting collector
which merges and applies these operations, we can describe in overview how the algo-
rithm operates.
The concurrent cycle collection algorithm is more complex than the synchronous
algorithm. As with other concurrent garbage collection algorithms, we must contend
with the fact that the object graph may be modified simultaneously with the collector
scanning it; but in addition the reference counts may be as much as a two epochs out of
date (because decrements are deferred by an epoch).
Our algorithm relies on the same basic premise as the synchronous algorithm:
namely, that given a subset of nodes, if deleting the internal edges between the nodes
in this subset reduces the reference count of every node in the subset to zero, then the
whole subset of nodes is cyclic garbage. (The subset may represent more than one in-
dependent cycles, but they are all garbage cycles.)
However, since the graph may be modified, we run into three basic difficulties.
Firstly, since we can not rely on being able to retrace the same graph, the repeated
traversal of the graph does not define the same set of nodes. Secondly, the deletion
of edges can disconnect portions of the graph, thus making the global test by graph
traversal difficult. Thirdly, reference counts may be out of date.
Our algorithm proceeds in two phases. In the first phase, we use a variant of the syn-
chronous algorithm as described in Section 3 to obtain a candidate set of garbage nodes.
We then wait until an epoch boundary and then perform the second phase in which we
test these to ensure that the candidates do indeed satisfy the criteria for garbage cycles.
The two phases can be viewed as enforcing a “liveness”and a “safety” property.The
first phase enforcesliveness by ensuring that potential garbage cycles are considered for
collection. The second phase ensures safety by preventing the collection of false cycles
induced by concurrent mutator activity.
4.2 Liveness: Finding Cycles to Collect
Essentially, we use the synchronous algorithm to find candidate cycles. However, due to
concurrent mutator activity, the graph may be changing and the algorithm may produce
incorrect results.

10 David F. Bacon and V.T. Rajan
To perform the concurrent cycle collection, we need a second reference count for
each object, denoted CRC(S). This is a hypothetical reference count which may be-
come incorrect due to concurrent mutator activity. In the implementation, we are able
to fit both reference counts, the color, and the buffered flag into a single header word by
using a hash table to hold count overflows, which occur very rarely.
The liveness phase of the concurrent algorithm proceeds in a similar manner to the
synchronous cycle collection algorithm, except that when an object is marked gray its
cyclic reference count (CRC) is initialized to its true reference count — the “true” ref-
erence count (RC) is not changed. Henceforward, the mark, scan, and collect phases
operate upon the cyclic reference count instead of the true reference count. In the
CollectWhite procedure, instead of collecting the white nodes as garbage, we color
them orange and add them to a set of possible garbage.
By using the cyclicreferencecount we ensurethatin the event of concurrentmutator
activity, the information about the true reference count of the objects is never lost.
In absence of mutator activity, the liveness phase will yield the set of garbagenodes,
and the safety phase will certify that this indeed is a set of garbage of nodes and we can
collect them.
However, the presence of concurrent mutator activity can cause live nodes to en-
ter the list in three different ways. Firstly, the mutator can add an edge, thus causing
the MarkGray procedure to incorrectly infer that there are no external edges to a live
object. Secondly, the mutator can delete an edge, thus causing scan procedure to in-
correctly infer a live object to be garbage. Thirdly, the deletion of edges concurrent to
running of the MarkGray and scan procedure can create gray and white nodes with
various values of cyclic reference counts. While eventually the reporting of the mutator
activity will cause these nodes to be detected and re-colored, if these nodes are encoun-
tered before they are re-colored they can mislead the runs of the above procedures into
inferring that they are garbage.
The output of phase one is a set of nodes believed to be garbage in the Cycle-
Buffer data structure. The CycleBuffer is divided into discrete connected com-
ponents, each of which forms a potential garbage cycle. Due to mutator activity, the
contents of the CycleBuffer can be a superset of the actual set of garbage nodes and
can contain some nodes that fail tests in the safety phase (this is discussed in detail in
Section 5).
4.3 Safety: Collecting Cycles Concurrently
The second (“safety”) phase of the algorithm takes as input a set of nodes and deter-
mines whether they form a garbage cycle. These nodes are marked with a special color,
orange, which is used to identify a candidate set in the concurrent cycle collector.
The safety phase of the algorithm consists of two tests we call the
-test and the
-test. If a subset of nodes of the object graph passes both the -test and the -test,
then we can be assured that the nodes in the subset are all garbage. Thus, correctness
of the safety phase of our algorithm is not determined by any property of the output of
the liveness phase which selects the subgraphs. This property of the safety phase of the
algorithm considerably simplifies the proof of correctness as well as modularizing the
code.

Concurrent Cycle Collection in Reference Counted Systems 11
In theory, it would be possible to build a cycle collector which simply passed ran-
dom sets of nodes to the safety phase, which would then either accept them as garbage
or reject them as live. However, such a collector would not be practical: if we indeed
pick a random subset of nodes from the object graph, the chances that they form a
complete garbage subgraph is very small. The job of the liveness phase can be seen as
finding likely sets of candidates for garbage cycles. If the mutator activity is small in a
given epoch, this would indeed be very likely to be true.
The -test consists of two parts: a preparation and an actual test. In the preparation
part, which is performed immediately after the candidate cycles have been found, we it-
erate over the subset and initialize the cyclic reference count of every node in the subset
to the reference count of the node. Then we iterate over every node in the subset again
and decrement the cyclic reference count of any children of the node that are also in the
subset. At the end of the preparation computation, the cyclic reference count of each
node in the subset represents the number of references to the node from nodes external
to the subset. In the actual test, which is performed after the next epoch boundary, we
iterate over every node in the subset and test if its cyclic reference count is zero.
If it is zero for every member of the set, then we know that there exists no reference
to this subset from any other node. Therefore, any candidate set that passes the
-
test is garbage, unless the reference count used during the running of the preparation
procedure is outdated due to an increment to one of the nodes in the subset.
This is ascertained by the
-test. We wait until the next epoch boundary, at which
point increment processing re-colors all non-black nodes and their reachable subgraphs
black. Then we scan the nodes in the candidate set and test whether their color is still
orange. If theyare all orange, we knowthat there has been no increment to the reference
count during the running of the preparation procedure and we say that the candidate set
passed the
-test.
Any subset of garbage nodes that does not have any external pointers to it will pass
both the tests. Note that we do not have to worry about concurrent decrements to the
members of the subset, since it is not possible for the reference count of any node to
drop below zero.
However, it is possible for a set of garbage to have pointers to it from other garbage
cycles. For example in Figure 1, only the candidate set consisting of the last node forms
isolated garbage cycle. The other cycles have pointers to them from the cycle to their
right.
We know that the garbage cycles in the cycle buffer cannot have any forward point-
ers to other garbage cycles (if they did, we would have followed them and included
them in a previous garbage cycle). Hence, we process the candidate cycles in the cycle
buffer in the reverse of the order in which we found them. This reasoning is described
more formally in Lemma 3 in Section 5.
When a candidate set passes both tests, and hence is determined to be garbage, then
we free the nodes in the cycle, which causes the reference counts of other nodes outside
of the cycle to be decremented. By the stability property of garbage, we can decrement
such reference counts without concern for concurrent mutation.
When we decrement a reference count to an orange node, we also decrement its
cyclic reference count (CRC). Therefore, when the next candidate cycle is considered

12 David F. Bacon and V.T. Rajan
(the previous cycle in the buffer), if it is garbage the
-test will succeed because we
have augmented the computation performed by the preparation procedure.
Hence when we reach a candidate set, the cyclic reference count does not include
the count of any pointers from a known garbage node. This ensures that all the nodes in
Figure 1 would be collected.
A formalism for understanding the structure of the graph in the presence of concur-
rent mutation, and a proof of correctness of the algorithm is presented in Section 5.
4.4 Pseudocode and Explanation
We now present the pseudocode with explanations for each procedure in the concurrent
cycle collection algorithm. The pseudocode is shown in Figures 4 and 5. The operation
of CollectCycles and its subsidiary procedures is very similar to the operation of
the synchronous algorithm of Figure 2, so for those procedures we will only focus on
the differences in the concurrent versions of the procedures.
Increment(S) The true reference count is incremented. Since the reference count
is being incremented, the node must be live, so any non-black objects reachable
from it are colored black by invoking ScanBlack. This has the effect of re-
blackening live nodes that were left gray or white when concurrent mutation in-
terrupted a previous cycle collection.
Decrement(S) At the high level, decrementing looks the same as with the syn-
chronous algorithm: if the count becomes zero, the object is released, otherwise it
is considered as a possible root.
PossibleRoot(S) For a possible root, we first perform ScanBlack. As with
Increment, this has the effect of re-blackening leftover gray or white nodes; it
may also change the color of some purple nodes reachable from S to black, but this
is not a problem since they will be considered when the cycle collector considers
S. The rest of PossibleRoot is the same as for the synchronous algorithm.
ProcessCycles() Invoked once per epoch after increment and decrement pro-
cessing due to the mutation buffers from the mutator threads has been completed.
First, FreeCycles attempts to free candidate cycles discovered during the pre-
vious epoch. Then CollectCycles collects new candidate cycles and Sigma-
Preparation prepares for the
-test to be run in the next epoch.
CollectCycles() As in the synchronous algorithm, three phases are invoked on
the candidate roots: marking, scanning, and collection.
MarkRoots() This procedure is the same as in the synchronous algorithm.
ScanRoots() This procedure is the same as in the synchronous algorithm.
CollectRoots() For each remaining root, if it is white a candidate cycle has been
discoveredstarting at that root.The CurrentCycle is initialized tobe empty, and
the CollectWhite procedure is invoked to gather the members of the cycle into
the CurrentCycle and color them orange. The collected cycle is then appended
to the CycleBuffer. If the root is not white, a candidate cycle was not found
from this root or it was already included in some previously collected candidate,
and the buffered flag is set to false. In either case, the root is removed from the
Roots buffer, so that at the end of this procedure the Roots buffer is empty.

Concurrent Cycle Collection in Reference Counted Systems 13
Increment(S) ScanRoots()
RC(S) = RC(S) + 1 for S in Roots
ScanBlack(S) Scan(S)
Decrement(S) CollectRoots()
RC(S) = RC(S) - 1 for S in Roots
if (RC(S) == 0) if (color(S) == white)
Release(S) CurrentCycle = empty
else CollectWhite(S)
PossibleRoot(S) append CurrentCycle
to CycleBuffer
Release(S) else
for T in children(S) buffered(S) = false
Decrement(T) remove S from Roots
color(S) = black
if (! buffered(S)) MarkGray(S)
Free(S) if (color(S) != gray)
color(S) = gray
PossibleRoot(S) CRC(S) = RC(S)
ScanBlack(S) for T in children(S)
color(S) = purple MarkGray(T)
if (! buffered(S)) else if (CRC(S) > 0)
buffered(S) = true CRC(S) = CRC(S) - 1
append S to Roots
Scan(S)
ProcessCycles() if (color(S) == gray
FreeCycles() and CRC(S) == 0)
CollectCycles() color(S) = white
SigmaPreparation() for T in children(S)
Scan(T)
CollectCycles() else
MarkRoots() ScanBlack(S)
ScanRoots()
CollectRoots() ScanBlack(S)
if (color(S) != black)
MarkRoots() color(S) = black
for S in Roots for T in children(S)
if (color(S) == purple ScanBlack(T)
and RC(S) > 0)
MarkGray(S) CollectWhite(S)
else if (color(S) == white)
remove S from Roots color(S) = orange
buffered(S) = false buffered(S) = true
if (RC(S) == 0) append S to CurrentCycle
Free(S) for T in children(S)
CollectWhite(T)
Fig.4. Concurrent Cycle Collection Algorithm (Part 1)

14 David F. Bacon and V.T. Rajan
SigmaPreparation() Refurbish(C)
for C in CycleBuffer first = true
for N in C for N in C
color(N) = red if ((first and
CRC(N) = RC(N) color(N)==orange) or
for N in C color(N)==purple)
for M in children(N) color(N) = purple
if (color(M) == red append N to Roots
and CRC(M) > 0) else
CRC(M) = CRC(M)-1 color(N) = black
for N in C buffered(N) = false
color(N) = orange first = false
FreeCycles() FreeCycle(C)
last = |CycleBuffer|-1 for N in C
for i = last to 0 by -1 color(N) = red
C = CycleBuffer[i] for N in C
if (DeltaTest(C) for M in children(N)
and SigmaTest(C)) CyclicDecrement(M)
FreeCycle(C) for N in C
else Free(N)
Refurbish(C)
clear CycleBuffer CyclicDecrement(M)
if (color(M) != red)
DeltaTest(C) if (color(M) == orange)
for N in C RC(M) = RC(M) - 1
if (color(N) != orange) CRC(M) = CRC(M) - 1
return false else
return true Decrement(M)
SigmaTest(C)
externRC = 0
for N in C
externRC = externRC+CRC(N)
return (externRC == 0)
Fig.5. Concurrent Cycle Collection Algorithm (Part 2)

Concurrent Cycle Collection in Reference Counted Systems 15
MarkGray(S) This is similar to the synchronous version of the procedure, with
adaptations to use the cyclic reference count (CRC) instead of the true reference
count (RC). If the color is not gray, it is set to gray and the CRC is copied from
the RC, and then MarkGray is invoked recursively on the children. If the color is
already gray,and if the CRC is not already zero, the CRC is decremented (the check
for non-zero is necessary because concurrent mutation could otherwise cause the
CRC to underflow).
Scan(S) As with MarkGray, simply an adaptation of the synchronous procedure
that uses the CRC. Nodes with zero CRC are colored white; non-black nodes with
CRC greater than zero are recursively re-colored black.
ScanBlack(S) Like the synchronous version of the procedure, but it does not need
to re-increment the true reference count because all reference count computations
were carried out on the CRC.
CollectWhite(S) This procedure recursively gathers white nodes identified as
members of a candidate garbage cycle into the CurrentCycle and colors them
orange as it goes. The buffered flag is also set true since a reference to the node
will be stored in the CycleBuffer when CurrentCycle is appended to it.
SigmaPreparation() After the candidate cycles have been collected into the
CycleBuffer, this procedure prepares for the execution of the
-test in the next
epoch. It operates individually on each candidate cycle C. First, each node S in C
has its CRC initialized to its RC and its color set to red. After this only the nodes of
C are red. Then for any pointer from one node in C to another node in C, the CRC
of the target node is decremented. Finally, the nodes in C are re-colored orange. At
the end of SigmaPreparation, the CRC field of each node S contains a count
of the number of references to S from outside of C.
FreeCycles() This procedureiteratesoverthe candidate cycles in the reverseorder
in which they were collected. It applies the safety tests (the
-test and the -
test) to each cycle and if it passes both tests then the cycle is freed; otherwise it is
refurbished, meaning that it may be reconsidered for collection in the next epoch.
DeltaTest(C) This procedure returns true if the color of all nodes in the cycle are
orange, which indicates that their have been no increments to any of the nodes in
the cycle.
SigmaTest(C) This procedure calculates the total number of external references
to nodes in the cycle, using the CRC fields computed by the SigmaPrepara-
tion procedure. It returns true if the number of external references is zero, false
otherwise.
Refurbish(C) If the candidate cycle has not been collected due to failing a safety
test, this procedure re-colors the nodes. If the first node in the candidate cycle
(which was the purple node from which the candidate was found) is still orange, or
if any node has become purple, then those nodes are colored purple and placed in
the Roots buffer. All other nodes are colored black and their buffered flags are
cleared.
FreeCycle(C) This procedure actually frees the members of a candidate cycle that
has passed the safety tests. First, the members of C are colored red; after this, only
the nodes in C are red. Then for each node S in C, CyclicDecrement decre-
ments reference counts in non-red nodes pointed to by S.

16 David F. Bacon and V.T. Rajan
CyclicDecrement(M) If a node is not red, then it either belongs to some other
candidate cycle or not. If it belongs to some other candidate cycle, then it is orange,
in which case both the RC and the CRC fields are decremented (the CRC field
is decremented to update the computation performed previously by the Sigma-
Preparation procedure to take the deletion of the cycle pointing to M into ac-
count). If it does not belong to some other candidate cycle, it will not be orange and
a normal Decrement operation is performed.
Forease of presentation,we havepresented the pseudocodein away that maximizes
readability. However, this means that as presented the code makes more passes over the
nodes than is strictly necessary. For instance, the first pass by SigmaPreparation
can be merged with CollectWhite, and the passes performed by DeltaTest and
SigmaTest can be combined. In the implementation, the passes are combined to min-
imize constant-factor overheads.
5 Proofs
In this section we prove the correctness of the concurrent cycle collection algorithm
presented in Section 4.
5.1 The Abstract Graph
For the purpose of the proof of correctness of the above tests for garbage it is useful
to define an abstract graph
for the epoch of the garbage collector. At beginning
of each epoch the collector thread gets a set of increments and decrements from each
of the mutator threads. If the increment refers to a new node, it implies the creation of
that node. In addition, each increment implies addition of a directed edge between two
nodes and each decrement implies deletion of an edge. The increments and decrements
do not provide the source of the edges, so in practice we cannot build this graph, nor do
we need to build it for the purpose of the algorithm. But for the purposes of the proof it
is useful to conceptualize this graph. The graph
denotes the graph that is generated
by adding nodes for each reference to a new node, and inserting and deleting edges to
corresponding to the increments and decrements at the beginning of epoch. In
addition, when a node is determined to be garbage and is freed, it is deleted from .
At the beginning of the first epoch we start with an empty graph
.
We can similarly define an abstract set of roots for each epoch . The roots are
either in the mutator stacks or in global (class static) variables. The roots in the mutator
stacks are named by the increments collected from the stack snapshots of each mutator
for the epoch. The roots from the global variables are the sources in edges implied by
increment and decrement operations whose source is a global variable instead of a heap
variable.
is simply the union of these two types of roots.
Given and , we can then define the set of garbage objects in , which we
denote
, as
that is, the set difference minus the transitive closure of the roots .

Concurrent Cycle Collection in Reference Counted Systems 17
5.2 Safety: Proof of Correctness
A garbage collector is safe if every object collected is indeed garbage. In this section
we prove the safety of our algorithm.
At the end of epoch
, the procedure ProcessCycles invokes FreeCycles
to collect the cycles identified as potential garbage during epoch .
Let the set denote the contents of the CycleBuffer generated during the cycle
collection in epoch
. This is the collection of orange nodes generated by the concurrent
variantof the synchronouscycle detection algorithm, which used the set ofpurple nodes
(denoted ) as roots to search for cyclic garbage.
is partitioned into the disjoint sets . Each is a candidate garbage
cycle computed by the cycle collection algorithm from a particular purple node in
.
Due to concurrent mutation, may contain nodes from .
Lemma 1. Any set
containing nodes that do not exist in (that is, )
will fail the -test.
Proof. The only way for new nodes to be added to in is by increment oper-
ations. However, all concurrent increment operations will have been processed before
we apply the
-test. Processing an increment operation invokes ScanBlack, so that
if the node in question is in , then that node (at least) will be re-colored from orange
to black. The presence of that black node in will cause the -test to fail.
Therefore, for a given epoch with , , and , let
denote the set containing the sets that passed the -test. The sets
in are denoted . Since all have passed the -test, all .
denote one of the sets , namely a set of nodes believed to be a garbage
cycle that has passed the -test.
denote a specific node in that collection ( ).
denote the reference count at a node in the epoch. By definition
RC(S) is the reference count of node S in graph .
denote the number of references to from nodes within .
denote the number of references to from nodes within as
determined by the
-test.
denote the hypothetical reference count for as computed by
the
-test.
denote , the complement of in .
Theorem 1. If
passes the -test, that is if we have computed the values of
for each as described in the procedure SigmaPreparation in Section 4 and
then is a set of garbage nodes ( ).

18 David F. Bacon and V.T. Rajan
Proof. From the above definitions, for every
,
Since we delay the processing of the decrements by one epoch, this ensures that the
following properties are true:
is non-negative and if it is zero, then is a garbage node.
is non-negative and if it is zero for every node in then is a
collection of garbage nodes.
During the
-test, we determine the number of references to node from nodes
within
. Therefore by definition,
The may differ from the because there may be new refer-
ences to from nodes within C that were added to , thereby increasing ;
or because there may be references to from nodes in that were deleted from ,
thereby decreasing
. If the collection passes the -test, then no references
were added to at any time during the last epoch.
Therefore,
and
If is zero from the -test, since is non-negative,it follows that
has to be zero too. Further, if for every node in the collection
, then the whole collection is garbage.
Lemma 1 and Theorem 1 show that the -test and the -test are sufficient to ensure
that any set of nodes that passes both tests is garbage. The following theorem shows
that both the tests are necessary to ensure this as well.
Theorem 2. Both
-test and -test are necessary to ensure that a candidate set
contains only garbage nodes.
Proof. We will prove by example. Consider the graph of nodes shown in Figure 6 (a).
The cycle was detected from the purple node
, which is the starting point from which
cycle collection is run. If the edge between nodes and is cut between the MarkGray
and the Scan routines,then the nodes and will be collected by the CollectWhite
routine and forma set
. These nodesare not garbage.However,since there havebeen
no increments to the reference counts of either of these nodes, this set will pass -test.
The decrements will be processed an epoch later, at epoch
, so the decrement
to node will not have an effect on the nodes and in the FreeCycles operation
performed in epoch . Even waiting for an additional epoch does not guarantee that
the fact that nodes
and will be detected by -test, since during epoch the edge

Concurrent Cycle Collection in Reference Counted Systems 19
1
0
2
0
(a)
a
b
e
d
c
2
0
1
0
f
g
1
0
h
(b)
2
1
2
1
cut
add
2
1
Fig.6. Race conditions uniquely detected (a) by the -test, and (b) by the -test. The purple
nodes from which cycle collection started were
and . Inside of each node is shown the refer-
ence count, or RC (top) and the cyclic reference count, or CRC (bottom).
from
to could be cut. Indeed, by making the chain of nodes arbitrarily long
and having a malicious mutator cut edges at just the right moment, we can have the
non-garbage set of nodes
pass the -test for arbitrarily many epochs. Hence the
-test alone cannot detect all live nodes in .
Now consider the graph of nodes shown in Figure 6 (b). The cycle is detected start-
ing with the purple node , from which cycle collection is run. If a new edge is added
from node to node before the MarkGray routine is run (shown as the dashed line
in the figure), the reference count of the node will be out of date. If the cycle col-
lector observes the newly added edge, the sum of the reference counts in
will
equal the sum of the edges. Hence the set of nodes will be collected by the
CollectWhite routine and form the set . If the increments are not processed be-
fore the
-test is done, then will pass the -test. Hence -test alone cannot detect
all live nodes in .
Notice that we are not claiming that the two race conditions shown in Figure 6 are
an exhaustive list of all possible race conditions that our algorithm will face. But these
two are sufficient to show the necessity of both the tests. Thus the two tests are both
necessary and sufficient to ensure the safety of the algorithm.
Finally we prove here the following Lemma that will be used in the next section,
since the proof uses the notation from the present section. We define a complete set of
nodes as one which is closed under transitive closure in the transpose graph; that is, a
complete set of nodes includes all of its parents.

20 David F. Bacon and V.T. Rajan
Lemma 2. If
is a complete set of nodes, then will pass both the -test and
the -test.
Proof. By the stability property of garbage, there can be no changes to the reference
counts of the nodes
, since . Therefore, passes the -test.
By the same reasoning,
Since is a complete set,
Therefore,
Hence, will pass the -test.
5.3 Liveness: Proof of Correctness
A garbage collector is live if it eventually collects all unreachable objects. Our concur-
rent algorithm is subject to some extremely rare race conditions, which may prevent the
collection of some garbage. Therefore, we provea weak liveness condition which holds
provided that the race condition does not occur in every epoch.
We can only demonstrate weak liveness because it is possible that a candidate cycle
contains a subset which is a complete garbage cycle, and some nodes in that
subset point to othernodes in which are live (see Figure 7). This is a result of our run-
ning a variant of the synchronous cycle collection algorithm while mutation continues,
thus allowing race conditions such as the ones shown in Figure 6 to cause Collect-
Cycles to occasionally place livenodes in a candidateset
. The resultingcandidate
set
will fail either the -test or the -test for that epoch. If this occurs, the cycle
will be reconsideredin the following epoch. Therefore,unless the race condition occurs
indefinitely, the garbage will eventually be collected.
Any garbage nodes that are not collected in epoch
are called the undiscovered
garbage of epoch .
In practice, we have been unable to induce the race condition that leads to undis-
covered garbage, even with adversary programs. However, this remains a theoretical
limitation of our approach.
We have solutions to this problem (for instance, breaking up a set that fails either of
the two tests into strongly connected components), but have not included them because
they complicate the algorithm and are not required in practice. We are also investigat-
ing another alternative, in which the entire cycle collection is based on performing a
strongly-connected component algorithm [4]; this alternative is also promising in that
the number of passes over the object graph is substantially reduced.

Concurrent Cycle Collection in Reference Counted Systems 21
B
A
P
i,k
D
C
P
i,k+1
P
i,k-1
B
i,j-1
B
i,j
Live
Data
Fig.7.Concurrent mutation of the nodes in D can cause candidate sets and to become
undiscovered garbage.
In this section we will prove that if a set of garbage nodes is free from the race
condition leading to undiscoveredgarbage, it will be collected in the epoch
; otherwise
it will be considered again in the epoch .
We previously defined
as the set of purple nodes in epoch , that is the set of
nodes from which the cycle detection algorithm begins searching for cyclic garbage in
CollectCycles.
Theorem 3. The purple set is maintained correctly by the concurrent cycle collection
algorithm: every garbage node is reachable from some purple node. That is,
Proof. The Decrement procedure ensures that the only garbage in the set is cyclic
garbage. In addition, Decrement adds all nodes having decrement to non-zero to the
purple set. Thus we know that any cyclic garbage generated during the processing of
increments and decrements in epoch
is reachable from the purple set. What remains
to be proved is that the purple set contains roots to any uncollected cyclic garbage from
the previous epochs.
We know this to be trivially true for epoch
. We assume that it is true for epoch
and prove that, after running the CollectCycles routine, it is true for epoch .
The result then follows by induction.
Let
be a member of the purple set in epoch that is the root of a garbage cycle.
The CollectRoots routine ensures that the root of each cycle generated by it is
stored in the first position in the bufferand takes all white nodes that are reachable from
it and unless it has been collected before (therefore reachable from another root node)
puts it in the current cycle. Since CollectCycles is a version of the synchronous

22 David F. Bacon and V.T. Rajan
garbage collection algorithm, and there can be no concurrent mutation to a subgraph
of garbage nodes, all such garbage nodes will be in the current cycle. In addition, any
other uncollected purple node reachable from
and the cycle associated with will be
added to the current cycle. If this latter purple node is garbage, then it will continue to
be reachable from
and hence proper handling of will ensure proper handling of
this node and its children.
The Refurbish routine will put this first node back into the purple root set unless:
a) the current cycle is determined to be garbage, in which case the entire cycle is freed
or b) the first node is determined to be live, in which case it is not the root of a garbage
cycle.
Hence the purpleset
will contain the roots of all the garbagecycles that survive
the cycle collect in epoch
.
Corollary 1. The cycle buffergenerated in the epochcontains all the garbagenodes
present in
. That is,
Proof. By Theorem 3, the root of everygarbage cycle is contained in . The procedure
CollectCycles is a version of the synchronous garbage collection algorithm. There
can be no concurrent mutation to a subgraph of garbage nodes. Therefore, all garbage
nodes will be in put into the cycle buffer
.
Unfortunately, due to concurrent mutation may contain some live nodes too. Let
denote a set of nodes that were collected starting from a root node .
If all the nodes in are live, then it will fail one of the two tests ( -test or -test)
and its root will be identified as a live node and discarded.
It is however possible that
contains a set of garbage nodes as well as a set
of live nodes as shown in Figure 7. There are three purple nodes .
CollectWhite processes
firstand creates the candidateset whichcon-
tains the nodes in A. Then CollectWhite processes and creates the candidate
set which contains the nodes in B, C, and D. This includes both the garbage nodes
in C reachable from another purple node not yet considered (
) and live nodes in
D.
In this case
will fail one of the two safety tests and the algorithm will fail to
detect that it contained some garbage nodes. Furthermore, the algorithm will fail to
detect any other garbage nodes that are pointed to by this garbage, such as
in
the figure. The roots
and will be put into , so the garbage cycles will
be considered again in epoch, and unless a malicious mutator is able to fool
CollectCycles again, it will be collected in that epoch. But the fact remains that it
will not be collected during the current epoch.
Let
be the set of nodes undiscovered garbage in epoch . That is, every member
of either has a live node in its cycle set or its cycle set is pointed to by a member of
. We will show that all the other garbage nodes (i.e. the set ) will be collected
in epoch .
Lemma 3. There are no edges from garbage nodes in
to where .

Concurrent Cycle Collection in Reference Counted Systems 23
Proof. The CollectRoots routine takes all white nodes that are reachable from the
root of the current cycle, colors them orange, and places them in a set
. Any nodes
that it reaches that were previously colored orange are not included in the set because
they havealready been included in some previous set
. Thus all nodes that are reach-
able from the current root exist in the current cycle or in a cycle collected previously.
Since the nodes in
were collected before the nodes in there can be no for-
ward pointers from the first to the second set, unless some edges were added after the
running of the CollectRoots routine. However, since there can be no mutation in-
volving garbage nodes, this is not possible.
Let be the set of all nodes collected by the procedure FreeCycles.
Theorem 4. All the garbage that is not undiscovered due to race conditions will be
collected in the
epoch. That is,
Proof. From Corollary 1 above, we know that every node in is contained in .
If a cycle
fails the -test, then we know that it has a live node. In that case, a
node in this cycle is either live, in which case it does not belong to , or it is a garbage
node that is undiscovered garbage, hence it belongs to . Thus none of these nodes
belong to
If a cycle fails the -test, it means that there is some undeleted edge from
outside the set of nodes in
. By Lemma 3, it cannot be from a garbage node that
comes earlier in the cycle buffer. If this is from a live node or from a garbage node
that is undiscovered garbage, then garbage nodes in this cycle, if any, belong to the set
. If it is not from an undiscovered garbage node, then that garbage node belongs to a
discovered garbage set later in the cycle buffer.
But in the FreeCycles routine we process the cycles in the reverse of the order
in which they were collected. As we free garbage cycles, we delete all edges from
the nodes in it. By Lemma 3, the last discovered garbage set cannot have any external
pointers to it. Therefore it will pass the
-test also. In addition, when we delete all
edges from this set, the next discovered garbage set will pass the -test. Hence, every
discovered garbage cycle set will pass both the tests.
Corollary 2. In the absence of the race condition leading to undiscovered garbage,
namely a mixture of live and garbage nodes in some set
, all garbage will be col-
lected. That is,
Proof. In this case, there are no live nodes in any that contains garbage and hence
is a null set. The result follows from Theorem 4.
6 Measurements
We now present measurements of the effectiveness of our concurrent cycle collection
algorithm within the Recycler, a reference counting garbage collector implemented as

24 David F. Bacon and V.T. Rajan
Program Description Applic. Threads Objects Percent
Size Allocated Acyclic
201 compress Compression 18 KB 1 0.2 M 73%
202 jess Java expert system 11 KB 1 17.4 M 19%
209 db Database 10 KB 1 6.6 M 10%
227 mtrt Multithreaded raytracer 571 KB 2 14.2 M 87%
228 jack Parser generator 131 KB 1 16.8 M 78%
portbob Business Object Benchmark 138 KB 3 7.9 M 61%
jalape˜no Jalape˜no compiler 1378 KB 1 19.2 M 7%
ggauss Cyclic torture test (synth.) 8 KB 1 32.5 M %
Table 2. Benchmarks and their overall characteristics.
partof, Jalape˜no, a Java virtual machine written in Java at theIBM T.J. Watson Research
Center.
The measurements in this section concentrate on the operation of the reference
counting system within the Recycler. In concurrently published work [3] we present
detailed measurements of the system as a whole, including a comprehensive perfor-
mance evaluation which shows that with sufficient resources, the Recycler achieves a
maximum 6 millisecond pause time without appreciably slowing downthe applications.
6.1 Benchmarks
Table 2 summarizes the benchmarks we used. Our benchmarks consist of a mixture of
SPEC benchmarks and other programs: portbob is an early version of the benchmark
recently accepted by SPEC underthe name jbb; jalape
˜
nois theJalape˜no optimizing
compiler compiling itself; and ggauss is a synthetic benchmark designed as a “torture
test” for the cycle collector: it does nothing but create cyclic garbage, using a Gaussian
distribution of neighbors to create a smooth distribution of random graphs.
Since we did no have source code for all benchmarks, application size is given as
the total class file size in kilobytes.
SPEC benchmarkswere run with “size100” for exactly two iterations, and the entire
run, including JIT time, was counted.
We ran the benchmarks with one more CPU than there are threads; the extra CPU
ran the concurrent collector.
The largest benchmark is jalape
˜
no, the Jalape˜no optimizing compiler compil-
ing itself. It allocates 19 million objects, of which only 8% are determined by the
classloader to be acyclic (and therefore marked green). The optimizer represents the
worst-case type of program likely to be seen by the cycle collector in practice: its data
structures consist almost entirely of graphs and doubly-linked lists.
6.2 Cycle Collection
Table 3 summarizes the operation of the concurrent cycle collection algorithm. Cycle
collection was performed every eight epochs or when the Roots buffer exceeded a

Concurrent Cycle Collection in Reference Counted Systems 25
Program Ep. Cyc. Roots Cycles Found Marked Refs. Trace/
Coll. Checked Coll. Gray White Orange Traced Alloc.
201 compress 40 9 12153 97 0 0 13020 726 484 62971 0.42
202 jess 127 33 155507 0 13 0 293995 14523 52 2281521 0.13
209 db 275 61 1261177 0 0 0 2793425 14551 0 23737713 3.57
227 mtrt 152 29 467334 10 0 2 3097618 1122418 654137 18558847 1.31
228 jack 230 47 151160 782 0 0 231356 75610 49141 875234 0.05
portbob 50 16 434071 0 0 5 678077 2279 5 6345488 0.80
jalape˜no 476 106 6382521 279790 0 0 7621940 3049754 2019731 49571627 2.58
ggauss 489 105 7111449 266666 0 0 7511620 6782726 6484782 37715868 1.16
Table 3. Cycle Collection. “Ep.” is the number of epochs; “Cyc. Coll.” is the number of cycle
collections. For “Cycles Found”, the number collected, and rejected due to the
- and -tests.
threshold size; usuallythe latter condition triggered cyclecollections sooner. They seem
to occur once every four to five epochs.
There were a number of surprising results. First of all, despite the large number of
roots considered, the number of garbage cycles found was usually quite low. Cyclic
garbage was significant in jalape
˜
no and our torture test, ggauss. It was also sig-
nificant in compress, although the numbers do not show it: multi-megabyte buffers
hang from cyclic data structures in compress, so the application runs out of memory
if its 97 cycles are not collected in a timely manner.
The Jalape˜no optimizing compiler and the synthetic graph generator both freed
about 20% of their objects with the cycle collector. We were surprised that the num-
ber was not higher, given that virtually all of the data structures were potentially cyclic.
However, it appears that even so, a large proportion of objects are simple (rather than
cyclic) garbage.
Of more than half a million candidate cycles found for the eight benchmarks, con-
currentmutation introducedonly 20 false candidate cycles, with most of the false cycles
being rejected by the
-test in jess. None of these rejected cycles was undiscovered
garbage that was collected later (that is, part of a set as described in Section 5).
In fact, we were unable to create false cycles artificially when we tried modifying
the ggauss program to turn it into a malicious mutator designed solely for the pur-
pose of fooling the CollectCycles algorithm. This demonstrates conclusively that
undiscovered garbage is a problem only in theory, and not in practice.
Table 3 also shows how the different phases of the cycle collection algorithm pro-
ceeded: marking (gray), looking like cyclic garbage (white), and provisionally identi-
fied as cyclic garbage (orange). The amount of marking varied widely according to the
benchmarks.
Finally, Table 3 shows the number of references that must be followed by the con-
current reference counting collector (“Refs. Traced”). We have also normalized this
against the total number of objects allocated (“Trace/Alloc”). The db benchmark re-
quired the most tracing per object. Apparently it performs far more modification of
its potentially cyclic data structures than other programs – presumably inserting and
removing database objects into its index data structure.

26 David F. Bacon and V.T. Rajan
Compared to tracing garbage collectors, the reference counting collector has an
advantage in that it only traces locally from potential roots, but has a disadvantage in
that the algorithm requires multiple passes over the subgraph. Furthermore, if the root
of a large data structure is entered into the root buffer frequently and high mutation
rates force frequent epoch boundaries, the same live data structure might be traversed
multiple times.
Our measurements show that a reference-counting based collector using our cycle
collection technique may perform very little tracing, or a large amount of tracing, and
that this is very application dependent.
Over all, the measurements presented here show that our cycle collection algorithm
is practical and capable of handling large programs, and in many cases should pro-
vide significantly increased locality and reduction in memory traffic over tracing-based
collectors.
The Recycler is described in greater detail and compared quantitatively to a parallel
mark-and-sweep collector by Bacon et al [3].
7 Related Work on Concurrent Collection
While numerous concurrent, multiprocessor collectors for general-purpose program-
ming languages have been described in the literature [8, 10, 11, 14, 15, 18, 19, 21, 26,
27], the number that have been implemented is quite small and of these, only a few
actually run on a multiprocessor [2, 8, 14, 11, 13, 24].
DeTreville’s work on garbage collectors for Modula-2+ on the DEC Firefly work-
station [8] is the only comparative evaluation of multiprocessor garbage collection tech-
niques. His algorithm is based on Rovner’s reference counting collector [26] backed by
a concurrent tracing collector for cyclic garbage. Unfortunately, despite having imple-
mented a great variety of collectors, he only provides a qualitative comparison. Nev-
ertheless, our findings agree with DeTreville’s in that he found reference counting to
be highly effective for a general-purpose programming language on a multiprocessor.
The Recycler differs in its use of cycle collection instead of a backup mark-and-sweep
collector.
Huelsbergen and Winterbottom [15] describe a concurrent algorithm (VCGC) that
is used in the Inferno system to back up a reference counting collector. They report
that reference counting collects 98% of data; our measurements for Java show that the
proportion of cyclic garbage is often small but varies greatly. The only measurements
provided for VCGC were on a uniprocessor for SML/NJ, so it is difficult to make mean-
ingful comparisons.
The only other concurrent, multiprocessor collector for Java that we know of is the
work of Domani et al [13, 12]. This is a generational collector based on the work of
Doligez et al [11], for which generations were shown to sometimes provide significant
improvements in throughput.
The other implemented concurrent multiprocessor collectors [2, 14, 11, 24] are all
tracing-based algorithms for concurrent variants of ML, and generally have signifi-
cantly longer maximum pause times than our collector. In addition, ML produces large
amounts of immutable data, thereby simplifying the collection process.

Concurrent Cycle Collection in Reference Counted Systems 27
The garbage collector of Huelsbergen and Larus [14] for ML achieved maximum
pause times of 20 ms in 1993,but onlyfortwo small benchmarks (Quicksortand Knuth-
Bendix). Their collector requires a read barrier for mutable objects that relies on pro-
cessor consistency to avoid locking objects while they are being forwarded. Read barri-
ers, even without synchronization instructions, are generally considered impractical for
imperative languages [17], and on weakly ordered multiprocessors their barrier would
require synchronization on every access to a mutable object, so it is not clear that the
algorithm is practical either for imperative languages or for the current generation of
multiprocessor machines.
Lins has presented a concurrent cycle collection algorithm [21] based on his syn-
chronous algorithm. Unlike the Recycler, Lins does not use a separate reference count
for the cycle collector; instead he relies on processor-supported asymmetric locking
primitives to prevent concurrent mutation to the graph. His scheme has, to our knowl-
edge, never been implemented. It does not appear to be practical on stock multiproces-
sor hardware because of the fine-grained locking required between the mutators and the
collector. Our algorithm avoids such fine-grained locking by using a second reference
count field when searching for cycles, and performing safety tests (the
-test and the
-test) to validate the cycles found.
Jones and Lins [16] present an algorithm for garbage collection in distributed sys-
tems that uses a variant of the lazy mark-scan algorithm for handling cycles. However,
they rely on much more heavy-weight synchronization (associated with message sends
and receives and global termination detection) than our algorithm. The algorithm has
never been implemented.
In terms of cycle collection systems that have been implemented, the closest to our
work is that of Rodrigues and Jones [25], who have implemented an algorithm for cycle
collection in distributed systems. However, they use a tracing collector for local cycles
and assume that inter-processor cycles are rare, and they use considerably more heavy-
weight mechanisms (such as lists of back-pointers) than we do; on the other hand they
also solve some problems that we do not address, like fault tolerance.
8 Conclusions
We have presented algorithms for the collection of cyclic data structures in reference
counted systems, starting with a synchronous algorithm which we then extended to han-
dle concurrent mutation without requiring any but the loosest synchronization between
mutator threads and the collector.
We presented detailed pseudocode and a proof of correctness of the concurrent al-
gorithm. We have implemented these algorithms as part of the Recycler, a concurrent
multiprocessor reference counting garbage collector for Java, and we presented mea-
surements that show the effectiveness of our algorithm over a suite of eight significant
Java benchmarks.
Our work is novel in two important respects: it represents the first practical use of
cycle collection in a reference counting garbage collector for a mainstream program-
ming language; and it requires no explicit synchronization between the mutator threads
or between the mutators and the collector.

28 David F. Bacon and V.T. Rajan
Another contribution of our work is our proof methodology, which allows us to rea-
son about an abstractgraph that neverexists in the machine, but is implied by the stream
of increment and decrement operations processed by the collector. In effect we are able
to reason about a consistent snapshot without ever having to take such a snapshot in the
implementation.
Our cycle collection algorithm forms a key part of the Recycler, a garbage col-
lector for Java, which achieves end-to-end execution times competitive with a parallel
mark-and-sweep collector while holding maximum application pause times to only six
milliseconds.
Acknowledgements
We thank Dick Attanasio, Han Lee, and Steve Smith for their contributions to the im-
plementation of the reference-counting garbage collector in which we implemented the
algorithms described in this paper, and the entire Jalape˜no team, without which this
work would not have been possible. We also thank the anonymous referees for their
comments which helped us to improve the paper.
References
[1] ALPERN, B., ET AL. Implementing Jalape˜no in Java. In OOPSLA’99 Conference Pro-
ceedings: Object-Oriented Programming Systems, Languages, and Applications (Denver,
Colorado, Oct. 1999). SIGPLAN Notices, 34, 10, 314–324.
[2] APPEL, A. W., ELLIS, J. R., AND LI, K. Real-time concurrent collection on stock multi-
processors. In Proceedings of theSIGPLAN Conference on Programming Language Design
and Implementation (Atlanta, Georgia, June 1988). SIGPLAN Notices, 23, 7 (July), 11–20.
[3] BACON, D. F., ATTANASIO, C. R., LEE, H. B., RAJAN, V. T., AND SMITH, S. Java
without the coffee breaks: A nonintrusive multiprocessor garbage collector. In Proceed-
ings of the SIGPLAN Conference on Programming Language Design and Implementation
(Snowbird, Utah, June 2001). SIGPLAN Notices, 36, 5 (May).
[4] BACON, D. F., KOLODNER, H., NATHANIEL, R., PETRANK, E., AND RAJAN, V. T.
Strongly-connected component algorithms for concurrent cycle collection. Tech. rep., IBM
T.J. Watson Research Center and IBM Haifa Scientific Center, Apr. 2001.
[5] BOBROW, D. G. Managing re-entrant structures using reference counts. ACM Trans.
Program. Lang. Syst. 2, 3 (July 1980), 269–273.
[6] CHRISTOPHER, T. W. Reference count garbage collection. Software – Practice and Expe-
rience 14, 6 (June 1984), 503–507.
[7] COLLINS, G. E. A method for overlapping and erasure of lists. Commun. ACM 3, 12 (Dec.
1960), 655–657.
[8] DETREVILLE, J. Experience with concurrent garbage collectors for Modula-2+. Tech.
Rep. 64, DEC Systems Research Center, Palo Alto, California, Aug. 1990.
[9] DEUTSCH, L. P., AND BOBROW, D. G. An efficient incremental automatic garbage col-
lector. Commun. ACM 19, 7 (July 1976), 522–526.
[10] DIJKSTRA, E. W., LAMPORT, L., MARTIN, A. J., SCHOLTEN, C. S., AND STEFFENS,
E. F. M. On-the-fly garbage collection: An exercise in cooperation. In Hierarchies and
Interfaces, F. L. Bauer et al., Eds., vol. 46 of Lecture Notes in Computer Science. Springer-
Verlag, New York, 1976, pp. 43–56.

Concurrent Cycle Collection in Reference Counted Systems 29
[11] DOLIGEZ, D., AND LEROY, X. A concurrent generational garbage collector for a multi-
threaded implementation of ML. In Conference Record of the Twentieth ACM Symposium
on Principles of Programming Languages (Charleston, South Carolina, Jan. 1993), ACM
Press, New York, New York, pp. 113–123.
[12] DOMANI, T., ET AL. Implementing an on-the-fly garbage collector for Java. In Proceed-
ings of the ACM SIGPLAN International Symposium on Memory Management (Minneapo-
lis, MN, Oct. 2000). SIGPLAN Notices, 36, 1 (Jan., 2001), 155–166.
[13] DOMANI, T., KOLODNER, E. K., AND PETRANK, E. A generational on-the-fly garbage
collector for Java. In Proceedings of the SIGPLAN Conference on Programming Language
Design and Implementation (June 2000). SIGPLAN Notices, 35, 6, 274–284.
[14] HUELSBERGEN, L., AND LARUS, J. R. A concurrent copying garbage collector for lan-
guages that distinguish (im)mutable data. In Proceedings of the Fourth ACM Symposium
on Principles and Practice of Parallel Programming (May 1993). SIGPLAN Notices, 28, 7
(July), 73–82.
[15] HUELSBERGEN, L., AND WINTERBOTTOM, P. Very concurrent mark-&-sweep garbage
collection without fine-grain synchronization. In Proceedings of the ACM SIGPLAN In-
ternational Symposium on Memory Management (Mar. 1999). SIGPLAN Notices, 34, 3,
166–174.
[16] JONES, R. E., AND LINS, R. D. Cyclic weighted reference counting without delay. In
PARLE’93 Parallel Architectures and Languages Europe (June 1993), A. Bode, M. Reeve,
and G. Wolf, Eds., vol. 694 of Lecture Notes in Computer Science, Springer-Verlag,
pp. 712–715.
[17] JONES, R. E., AND LINS, R. D. Garbage Collection. John Wiley and Sons, 1996.
[18] KUNG, H. T., AND SONG, S. W. An efficient parallel garbage collection system and its
correctness proof. In IEEE Symposium on Foundations of Computer Science (1977), IEEE
Press, New York, New York, pp. 120–131.
[19] LAMPORT, L. Garbage collection with multiple processes: an exercise in parallelism. In
Proceedings of the 1976 International Conference on Parallel Processing (1976), pp. 50–
54.
[20] LINS, R. D. Cyclic reference counting with lazy mark-scan. Inf. Process. Lett. 44, 4 (Dec.
1992), 215–220.
[21] LINS, R. D. A multi-processor shared memory architecture for parallel cyclic reference
counting. Microprocessing and Microprogramming 35, 1–5 (Sept. 1992), 563–568. Pro-
ceedings of the 18th EUROMICRO Conference (Paris, France).
[22] MART
´
INEZ, A. D., WACHENCHAUZER, R., AND LINS, R. D. Cyclic reference counting
with local mark-scan. Inf. Process. Lett. 34, 1 (1990), 31–35.
[23] MCCARTHY, J. Recursive functions of symbolic expressions and their computation by
machine. Commun. ACM 3 (1960), 184–195.
[24] NETTLES, S., AND O’TOOLE, J. Real-time garbage collection. In Proceedings of the SIG-
PLAN Conference on Programming Language Design and Implementation (Albuquerque,
New Mexico, June 1993). SIGPLAN Notices, 28, 6, 217–226.
[25] RODRIGUES, H. C. C. D., AND JONES, R. E. Cyclic distributed garbage collection with
group merger. In Proceedings of the Twelfth European Conference on Object-Oriented
Programming (Brussels, July 1998), E. Jul, Ed., vol. 1445 of Lecture Notes in Computer
Science, Springer-Verlag, pp. 249–273.
[26] ROVNER, P. On adding garbage collection and runtime types to astrongly-typed, statically-
checked, concurrent language. Tech. Rep. CSL–84–7, Xerox Palo Alto Research Center,
July 1985.
[27] STEELE, G. L. Multiprocessing compactifying garbage collection. Commun. ACM 18, 9
(Sept. 1975), 495–508.