TECHNOLOGY WHITE PAPER

Understanding Java
Garbage Collection
And what you can do about it
 Understanding Java Garbage Collection

collection behavior, as well as examples of how certain

choices can result in impressive application behavior.

Introduction
The Java programming language utilizes a managed
runtime (the Java Virtual Machine, or JVM) to improve
developer productivity and provide cross-platform
portability. Because different operating systems and
hardware platforms vary in the ways that they manage
memory, the JVM performs this function for the develop-
er, allocating memory as objects are created and freeing
it when they are no longer used. This process of freeing
unused memory is called ‘garbage collection’ (GC), and
is performed by the JVM on the memory heap during
application execution.
This paper explains in more detail how garbage
collection works, the different algorithm types employed Java garbage collection can have a big impact on
application performance and throughput. As the JVM
by commercially available JVMs, and how developers
heap size grows, so does the amount of time that an
and architects can make better informed decisions application must pause to allow the JVM to perform GC.
The result can be long, unexpected pauses that can
delay transactions, deteriorate application throughput,
cause user-session time-outs, force nodes to fall out of
clusters, or result in even more severe business related
Executive Summary losses (e.g. drop in revenue or damage to reputation).
Garbage Collection (GC) is an integral part of application
behavior on Java platforms, yet it is often misunderstood.
This paper explains in more detail how garbage
Java developers need to understand how GC works and
collection works, the different algorithm types employed
how the actions they can take in selecting and tuning
by commercially available JVMs, and how developers
collector mechanisms, as well as in application architec-
and architects can make better informed decisions on
ture choices, can affect runtime performance, scalability
which garbage collector to use and how to maximize
and reliability.
application performance.
This white paper reviews and classifies the various
Why Care About the Java Garbage Collector?
garbage collectors and collection techniques available in
Overall garbage collection is much better and more
JVMs today. This paper provides an overview of common
efficient than you might think. It’s much faster than
garbage collection techniques, algorithms and defines
malloc() at allocating memory and dead objects cost
terms and metrics common to all collectors including:
nothing to collect (really!). GC will find all the dead
• Generational
objects, even in cyclic graphs, without any assistance
• Parallel
from the developer. But in many ways garbage collection
• Stop-the-world
is much more insidious than many developers and
• Incremental
architects realize.
• Concurrent
• Mostly-concurrent
For most collectors GC related pauses are proportional
to size of their heaps which is approximately 1 second
The paper classifies each major JVM collector’s mecha-
for each gigabyte of live objects. So, a larger heap
nisms and characteristics and discusses the trade-offs
(which can be advantageous for most apps) means a
involved in balancing requirements for responsiveness,
longer pause. Worse yet, if you run a 20 minute test
throughput, space, and available memory across varying
and tune until all the pauses go away, the likelihood is
scale levels. The paper concludes with some pitfalls,
that you’ve simply moved the pause to the 21st minute.
common misconceptions, and “myths” around garbage

TABLE OF CONTENTS

Why Care About the Java Garbage Collector? 2 Remembered Set 5 Oracle HotSpot CMS 7
Classifying the Collector 3 Commercial Implementations 6 Oracle HotSpot G1GC (Garbage First) 7
Steps in Garbage Collection 3 What Developers and Architects Can Do 6 Azul Systems Zing C4 8
Mark 3 Garbage Collection Metrics 6 GC Tuning Observations 9
Sweep 4 The Need for Empty Memory for GC 6 Summary 10
Compact 4 GC Strategy: Delaying the Inevitable 6 About Azul Systems 10
Types of Collectors 5 Choosing a Garbage Collector 7 Appendix A: Garbage Collection Terminology 10
2 Generational Collectors 5 Oracle HotSpot ParallelGC 7
 Understanding Java Garbage Collection

So unfortunately, the pause will still happen and your Concurrent collector – performs garbage collection
application will suffer. Inaddition, the presence of concurrently while application execution continues.
garbage collection doesn’t eliminate object leaks –
Parallel collector – uses multiple threads. A collector
the developer still has to find and fix references holding
can be concurrent but not parallel, and it can be
those leaked objects.
concurrent AND parallel. (Side note – be cautious when
researching older literature on garbage collection, since
The good news is Java does provide some level of GC
what we used to call parallel is now called concurrent.)
control. Developers and architects can make decisions
that can adjust application performance, due to the Stop-the-world (STW) – is the opposite of concurrent.
behavior of the garbage collector. For example, in C++ it It performs garbage collection while the application is
makes sense to null every reference field when it’s no completely stopped.
longer needed. However, in a Java program, coding in
Incremental – performs garbage collection as a series
nullifiers everywhere is disastrous and far worse than
of smaller increments with potentially long gaps in
coding in nothing. If every single class uses a finalizer to
between. The application is stopped during garbage
null reference fields, the garbage collector will potentially
collection but runs in between increments.
have to perform millions of object finalizations per GC
cycle – leading to very long garbage collection pauses. Moving – the collector moves objects during garbage
collection and has to update references to those live
objects.
Conservative – most non-managed runtimes are
conservative. In this model, the collector is unsure of
whether a field is a reference or not, so it assumes that
it is. This is in contrast to a Precise Collector.
Precise – a precise collector knows exactly where every
possible object reference is. A collector cannot be a
moving collector without also being precise, because you
have to know which references to fix when you move the
live objects. Precise collectors identify the live objects in
the memory heap, reclaim resources held by dead
objects and periodically relocate live objects.

Trying to solve garbage collection at the application Most of the work the virtual machine does to be precise,
programming layer is dangerous. It takes a lot of practice is actually in the compiler, not the collector itself. All
and understanding to get it right; time that could better commercial JVMs today are moving and precise.
spent building value-added features. And, even if you
make all the right decisions, it is likely that other code
your application leverages will not be optimized or the Safepoint opportunities in your code should be
application workload changes over time, and your frequent. If the garbage collector has to wait for a
application will still have GC related performance issues. safepoint that is minutes (or longer) away, your

Also, depending on the characteristics of your applica- application could run out of memory and crash
tion, choosing the wrong garbage collector type or using before garbage can be collected.
the wrong settings can greatly increase pause times or
even cause out-of-memory crashes. With a proper
understand of garbage collection and what your available Steps in Garbage Collection
options are, you can make better informed decisions and Before the garbage collector can reclaim memory, it
product choices that can improve the performance and must ensure the application is at a ‘GC safepoint’. A
reliability of your application at runtime. GC safepoint is a point or range in a thread’s execution
where the collector can identify all the references in the
Classifying the Collector thread’s execution stack. The terms ‘safepoint’ and ‘GC
Garbage collectors are divided into several types. For safepoint’ are often used interchangeably, however many
each type some collectors are categorized as ‘mostly’, types of safepoints exist, some of which require more
as in ‘mostly concurrent’. This means that sometimes it information than a GC safepoint. A ‘Global Safepoint’ is
doesn’t operate according to that classification and has when all application threads are at a safepoint.
a fallback mechanism for when that occurs. So, a
‘mostly concurrent’ collector may operate concurrently Safepoint opportunities in your code should be frequent.
with application execution and only occasionally If the garbage collector has to wait for a safepoint that is
stop-the-world if needed. minutes (or longer) away, your application could run out
3
 Understanding Java Garbage Collection

of memory and crash before garbage can be collected.

Once the GC safepoint is reached, garbage collection
can begin.

Mark
This phase, also known as ‘trace’, finds all the live
objects in the heap. The process starts from the ‘roots’,
which includes thread stacks, static variables, special
references from JNI code and other areas where live
objects are likely to be found. A reference to an object
can only prevent the object from being garbage collected,
if the reference chains from a GC root.

The garbage collector ‘paints’ any objects it can reach as

‘live’. Any objects left at the end of this step are ‘dead’. If
any objects are still reachable that the developer thought
were dead, it’s an object leak, a form of memory leak.

The work of a marker is linear to the amount of live

objects and references, regardless of the size of the
objects. In other words it takes the marker the same Compact
amount of time to paint 1,000 10KB objects as 1,000 Over time, the Java memory heap gets ‘fragmented’,
1MB objects. where the dead space between objects is no longer
large enough to hold new objects, making new object
In concurrent marking all reachable objects are being allocation slower. If your application creates objects of
marked as live, but the object graph is mutating (i.e. variable sizes, fragmentation will happen more quickly.
changing) as the marker works. This can lead to a XML is a great example of this. The format is defined but
classic concurrent marking the size of the information in the object is not controlled,
race. The application can move often leading to objects with great variations in sizes and
Over time, the Java memory heap a reference that has not yet a fragmented heap.
gets ‘fragmented’, where the been seen by the marker into
an object that has already been In the Compact phase the garbage collector relocates
dead space between objects is visited. If this change is not live objects to free up contiguous empty space. As these
no longer large enough to hold intercepted or prevented in objects are moved, the collector must fix all references
new objects, making new object some way, it can corrupt the in the threads to these live objects, called ‘remapping’.
heap. The object would be Remap has to cover all references that could point to an
allocation slower. collected, even though a object, so it usually scans everything. The amount of
reference to it still exists. work done in this phase is generally linear to the size of
the live set.
Usually a ‘write barrier’ is used to prevent this condition.
The write barrier captures changes to object references Incremental compaction is used in a couple of
(e.g. in a card table) that might otherwise be missed by commercial collectors (Oracle G1 and the Balanced
the marker. With this information, the marker can revisit Collector from IBM). This technique assumes that some
all mutated references and track new mutations. When regions of memory are more popular than others,
the set is small enough, a stop-the-world pause can be although this is not always the case depending on the
used to catch up, making the collector ‘mostly’ concur- application. The GC algorithm tracks cross-region
rent. But it is important to note that the collector is remembered sets (i.e. which region points to which).
sensitive to the mutation rate and the amount of work This allows the collector to compact a single region at
done grows with the mutation rate and may fail to finish. a time and only scan regions pointing into it when
remapping all potential references. The collector
Sweep identifies region sets that fit into limited pause times,
In this phase the garbage collector scans through the allowing the maximum time for application interruption
heap to identify the locations of ‘dead’ objects and to be controlled. Large heaps have fewer non-popular
tracks their location, usually in some sort of ‘free list’. regions; the number of regions pointing into a single
Unlike the Mark phase, the work done by the Sweep region tends to be linear to the size of the heap.
phase is linear to the size of the heap, not the size of Because of this, the work for this type of compaction
the live set. If your application is using a very large heap can grow with the square of the heap size.
with very little left alive, Sweep still has to scan on the
entire heap.
4
 Understanding Java Garbage Collection

Types of Collectors
Mark/Sweep/Compact Collector – performs the three
phases as three separate steps.

Mark/Compact Collector – skips the sweep and moves

live objects to a contiguous area of the heap.
Copying Collector – performs all three phases in one
pass. A copying collector is pretty aggressive. It uses a
‘from’ and ‘to’ space and moves all the live objects then
fixes the references all in one pass. When the ‘from’
space is empty the collection is complete. Work done
in a copying collector is linear to the size of the live set.

Fig. 1: Comparing collector types

Mark/Sweep/Compact Mark/Compact Copying to place live objects and the
work done is linear to the size
Amount of memory 1x the size of the 2x the size of 2x the size of of the live set which is small.
needed to perform live set plus a little the live set the live set A generational collector doesn’t
collection more need memory equal to 2x the live
Monolithic – the whole No No Typically yes set, since objects can spillover
heap must be garbage into the old generation space.
collected at once This compensates for the main
downside of a copying collector.
Amount of work Size of heap Size of live set Size of live set You can allow the young
linear to (in sweep)
generation to get completely
Requires large No Yes Yes full before collecting it.
amount of ‘free’
memory Deciding when to promote
objects can dramatically improve
Fastest for ‘Full’ heaps with little Heaps that become Low live object
free memory; large fragmented in counts efficiency. Keeping objects in the
heaps M/S/C young generation a little longer
may allow many of them to die
Fragments the Yes Yes No and save collection time. If you
heap keep them too long the young
generation can run out of
Generational Collectors space or ruin the generational
A generational collector is based on the hypothesis that assumption altogether. Waiting too long to promote can
most objects die young. The application creates them, also dramatically increase the work needed to copy the
but quickly doesn’t need them anymore. Often a method live objects and therefore the time it takes to do GC.
creates many objects but never stores them in a field.
When the method exits those objects are ready for Remembered Set
garbage collection. The developer can set a ‘generational Generational collectors use a ‘remembered set’ to
filter’ that reduces the rate of allocation into the old track all references into the young generation from
generation. This filter is the only way to keep up with the outside, so the collector doesn’t have to scan for
today’s CPU throughput – otherwise applications can them. This set is also used as part of the ‘roots’ for
create objects much faster than garbage collection can the garbage collector. A common technique is ‘card
clean them up. marking’, which uses a bit (or byte) indicating that a
word or region in the old generation is suspect. These
For applications where this hypothesis holds true, it ‘marks’ can be precise or imprecise, meaning it may
makes sense to focus garbage collection efforts on the record the exact location or just a region in memory.
‘young generation’, and promote objects that live long Write barriers are used to track references from the
enough to an ‘old generation’ which can be garbage young generation into the old generation and keep
collected much less frequently as it fills up. the remembered set up-to-date.

Because these young generation objects die quickly, Oracle’s HotSpot uses what’s called a ‘blind store’.
the live set in the young generation takes up a small Every time you store a reference it marks a card. This
percentage of the available space. Thus a moving works well, because checking the reference takes more
collector makes sense, since we have space in which CPU time, so the system saves time by just marking
5 the card.
 Understanding Java Garbage Collection

Commercial Implementations objects, and is only relevant for a Mark/Compact

Commercial server-side JVMs typically use a copying collector. For Mark/Sweep/Compact, sweep time is also
collector for the young generation that employs a important, which indicates how long it takes the collector
monolithic, stop-the-world collection. In other words the to locate all the dead objects. Cycle time for the collector
collector stops application processing and copies the is the total time from the start of garbage collection,
entire live set into a new section of the heap. The old until memory is freed and available for use by the
generation usually uses a Mark/Sweep/Compact application.
collector which may be stop-the-world, concurrent,
mostly concurrent, incremental stop-the-world or The Need for Empty Memory for GC
mostly incremental stop-the-world. Garbage collectors need at least some amount of empty
memory in order to work. More empty memory makes it
1
Normally, the system wants to be able to get to the easier (and faster) for the garbage collector. Doubling
large live set in the old generation without having to empty memory halves the work done by the collector and
stop at some increment. halves the CPU consumption needed to run. This is often
1 Translation, in order: it may stop the application entirely to perform the the best tool for efficiency.
collection, perform collection concurrent with application processing, collect
concurrently with the application up to a point when it (for example) gets
behind, and has to stop the application to catch up, stop the application
To illustrate, here are a couple of intuitive limits. If we
processing for shorter periods to do part of the garbage collection at a time,
have infinite empty memory we would never have to
|or it may do these incremental collections for as long as possible, before it
has to stop the application to complete GC.
collect and GC would never use any CPU time. If we have
exactly 1 byte of empty memory at
all times, the collector would have
to work very hard and GC would
Garbage collectors need at least use up 100% of CPU time.
Overall, garbage collection CPU
some amount of empty memory in
time follows an approximate 1/x
order to work. More empty memory curve between these two limits,
with effort dropping as empty
makes it easier (and faster) for the
memory increases.
garbage collector.
Mark/Compact and Copying
collector work is linear to the size
of the live set. How often each
needs to run is determined by
the amount of empty memory. Since collection is a fixed
amount of work each time, doing this work less often
What Developers and Architects
is more efficient. In these two types of collectors the
Can Do
amount of empty memory available doesn’t control
First, understand the characteristics of your application
the length of the garbage collection pause, only the
and the basics of how garbage collection works.
frequency. On the other hand Mark/Sweep/Compact
work grows as the heap grows. More empty memory for
Garbage Collection Metrics
a collector that pauses for sweeping, means less
Many characteristics of your application will affect
frequent but longer pauses.
garbage collection and performance at runtime. First is
how fast your application is allocating objects in memory,
Now that we understand what the characteristics are for
or the allocation rate. Next is how long those objects
our application, we can make changes that will improve
live. Do you have a fairly typical application where
performance, scalability and reliability.
objects die young, or do you have many objects that
are needed for a long time? Your program may also be
GC Strategy: Delaying the Inevitable
updating references in memory, called the mutation rate.
Although compaction is unavoidable in practice, many
The mutation rate is generally linear to the amount of
GC tuning techniques focus on delaying a full GC as
work the application is doing. And finally, as objects are
long as possible and freeing the ‘easy’ empty space
created and are dying, another set of metrics to track is
as quickly as possible.
the live set (also called the ‘heap population’) and the
heap shape, which is the shape of the live object graph.
Generational garbage collection can be partially
effective at delaying the inevitable. Young generation
The mark time and compaction time are the most
objects are collected frequently and this doesn’t take
important metrics to track for overall garbage collection
much time. But eventually, space in the old generation
cycle time. Mark time is how long it takes for the
must be reclaimed using a monolithic, stop-the-world
collector to find all live objects on the heap. Compaction
collection. Another delay strategy is to perform
time is how long it takes to free up memory by relocating
6
 Understanding Java Garbage Collection

Oracle’s HotSpot ParallelGC

This is the default collector for HotSpot. It uses a
monolithic, stop-the-world copying collector for the young
generation and a monolithic, stop-the-world Mark/
Sweep/Compact algorithm for the old generation.

Oracle’s HotSpot CMS

The Concurrent Mark/Sweep collector (CMS) is an option
in HotSpot. It attempts to reduce the old generation
pauses as much as possible by concurrently marking
and sweeping the old generation without compacting.
Once the old generation becomes too fragmented, it
falls back to a monolithic, stop-the-world compaction.
concurrent marking and sweeping, but skip compaction.
Freed memory can be tracked in lists and reused without CMS performs mostly concurrent marking. It marks
moving live objects. But over time this will leads to concurrently while the application is running and tracks
fragmentation, forcing a compaction. any changes (mutations) in card marks which are
revisited as needed. CMS also does concurrent sweep-
Finally, some collectors rely on the idea that much of ing, which frees up space occupied by dead objects and
the heap is not popular. A non-popular region will only creates a free list for allocation of new objects. When
be pointed to from a small portion of the overall heap. objects can no longer be promoted to the old generation
Compaction can be done on non-popular regions or concurrent marking fails, the system will throw a
incrementally with short stop-the-world pauses to free ‘promotion failure’ message and/or pause for a second
up space. However, at some point the popular regions or more while it performs compaction.
will need to be compacted, causing an application
pause. The young generation uses a monolithic, stop-the-world
copying collector, just like the ParallelGC option.
The bottom line is that despite numerous techniques
and creative ways to tune away garbage collection Oracle’s HotSpot G1GC (Garbage First)
pauses, competition is inevitable with most commercial G1 is an option in HotSpot. Its goal is to avoid,
collectors. Developers and architects need to make good “as much as possible” a full GC. G1 uses a mostly
decisions about which collector to use to maximize concurrent marker for the old generation. It marks
application performance. concurrently as much as possible, then uses a

Choosing a Garbage Collector

The following table summarizes the types of popular commercially available garbage collectors.

Collector Name Young Generation Old Generation

Oracle’s HotSpot ParallelGC Monolithic, stop-the-world, Monolithic, stop-the-world,
copying Mark/Sweep/Compact

Oracle’s HotSpot CMS (Concurrent Monolithic, stop-the-world, Mostly concurrent, non-compacting

Mark/Sweep) copying

Oracle’s HotSpot G1 (Garbage First) Monolithic, stop-the-world, Mostly concurrent marker,

copying mostly incremental compaction,
fall back to monolithic stop the world

Oracle’s JRockit* Dynamic Garbage Monolithic, stop-the-world, Mark/Sweep - can choose mostly
Collector copying concurrent or parallel, incremental compaction,
fall back to monolithic stop-the-world

IBM J9* Balanced Monolithic, stop-the-world, Mostly concurrent marker, mostly incremental
copying compaction, fall back to monolithic stop-the-world

IBM J9* Opt throughput Monolithic, stop-the-world, Parallel Mark/Sweep, stop-the-world compaction
copying

Zing C4 (Continuously Concurrent, compacting Concurrent, compacting

Concurrent Compacting
Collector)

*Can choose a single or 2-generation collector

7
 Understanding Java Garbage Collection

stop-the-world pause to catch up on mutations and

reference processing. G1 tracks inter-regional relation-
ships in remembered sets and does not use fine-grained
free lists. It uses stop-the-world, mostly incremental
compaction and delays compaction of popular objects
and regions as much as possible. G1 falls back to
monolithic, stop-the-world full compaction of these
popular areas when needed.

The young generation uses a monolithic, stop-the-world

copying collector, just like the ParallelGC and CMS
options.

Azul Systems Zing C4

C4 (Continuously Concurrent Compacting Collector) pass concurrent marker. No matter how fast your
is the default in Azul Systems’ Zing. C4 concurrently application mutates the heap, C4 can keep up. The C4
compacts both the young generation and old generation. collector also performs concurrent reference processing
Unlike other algorithms, it is not ‘mostly’ concurrent, but (including weak, soft and final references) and relocation
fully concurrent, so it never falls back to a stop-the-world and remapping are both concurrent.
compaction.
C4 also uses ‘quick release’ to make freed memory
C4 uses a Load Value Barrier (LVB) to verify each heap available quickly to the application and for the relocation
reference as correct when loaded. Any heap references of objects. This enables ‘hand over hand’ compaction
that point to relocated objects are caught and fixed in that does not require empty memory to function.
this self-healing barrier. C4 has a guaranteed single

Test Raw Bench Throughput of Collectors

140000

120000
Te s t B e n c h o p s / s e c

100000

80000
CMS
60000
parrallel-gc
40000
C4-NoConcYoung
20000
C4
0
0 5 10 15 20 25 30 35
Heap Size (GB)

Worst Case Response Times

Worst case response time [sec]

10

1 CMS

parrallel-gc

0.1 C4-NoConcYoung

C4

0.01

0 5 10 15 20 25 30 35

8 Heap Size (GB)

 Understanding Java Garbage Collection

With most collectors no ‘one size fits all’

answer exists. Developers and architects have
to tune garbage collection carefully and retune
every time the application, environment or
anticipated load changes.

However, performance of an application on the Azul

Zing C4 collector is insensitive to the ‘usual’ tuning
parameters. Because it marks and compacts
GC Tuning Observations concurrently in both the young and old generations
Garbage collection tuning for most collectors is hard to the only important parameter is heap size.
get right, even when you understand the characteristics
of your application. The figure below shows two sets of When you set the heap size in the Zing runtime, the C4
tuning parameters for the CMS collector in the HotSpot collector computes the GC time it needs in order to keep
JVM. While they may use similar parameter, they are very up with allocation rates. You do not need to set ratios or
different and in some areas diametrically opposed. evacuation times. Zing’s C4 GC will fire when needed
Yet the performance of your application could be using threads that are separate from your application
optimized with either set, depending on its particular threads. This allows worst case pause times to drop by
characteristics. With most collectors no ‘one size fits all’ orders of magnitude compared to other GC strategies.
answer exists. Developers and architects have to tune
garbage collection carefully and retune every time the With GC out of the way, time to safepoint emerges as
application, environment or anticipated load changes. the next dominant source of delays, and even lower
Getting these parameters wrong can cause latencies may be achieved using other controls, typically
unexpected, long pauses during peak load times. in the OS, or in the JVM through JIT compiler adjustment.

Sample HotSpot JVM CMS Settings

Sample Azul Zing C4 Collector Settings

9
 Understanding Java Garbage Collection

Appendix A: Garbage Collection Terminology

Compaction The garbage collection phase that defrag-

ments the heap, moves objects in memory, remaps all
affected references and frees contiguous memory
regions.

Concurrent A type of garbage collection algorithm that

where GC is done while the application is running.

Copying A garbage collector that copies that performs

mark/sweep/compact all at once by copying live objects
to a new area in memory.
Summary Dead object An object that is no longer being referenced
Garbage Collection (GC) is an integral part of by the application.
application behavior on Java platforms, yet it is often
misunderstood. Java developers can improve application GC safepoint A point or range in a thread’s execution
performance, scalability and reliability by understanding where the collector can identify all the references in the
how GC works and by making better garbage collector thread’s execution stack.
choices.
Generational Objects in memory are split between a
The main drawback of most garbage collectors is the young generation and old generation and garbage
collected separately.
need for long application pauses. These pauses are the
result of an inevitable requirement to compact the heap
Incremental Garbage collects only a portion of the heap
to free up space. Collectors use different strategies to at a time.
delay these events, but compaction is inevitable for all
commercial available collectors except Azul C4, which Live object One that is still being referenced by the
employs a Continuously Concurrent Compacting application.
Collector that avoids pauses altogether.
Marking The garbage collection phase that identifies all
About Azul Systems live objects in the heap.
Azul Systems delivers Java solutions that meet the
needs of today’s real-time business. Designed and Monolithic Garbage collects the entire heap at once.
Mutator Your application, which is changing (mutating)
optimized for x86 servers and enterprise-class
references in memory.
workloads, Azul’s Zing is the only Java runtime that
supports highly consistent and pauseless execution Mutator Your application, which is changing (mutating)
for throughput-intensive and QoS-sensitive Java references in memory.
applications. Azul’s enhanced Java technologies also
enable organizations optimize the value of their capital Parallel A collector that uses multiple threads.
investments while ensuring that their human capital is
free to generate new features, unlocking creativity and Pause Time period when the application is stopped while
driving competitive advantage. garbage collection is occurring.

Precise A precise collector knows exactly where every

possible object reference is.
Contact Azul Promotion Allocating an object from the young generation
To discover how Zing can improve scalability, to the old generation of the heap.
consistency and performance of all your
Java deployments, contact: Remembered Set Tracks all references into the young
generation from the outside so the collector doesn’t have
Azul Systems, Inc. to scan for them.
385 Moffett Park Drive, Suite 115
Sunnyvale, CA 94089 USA Roots Starting points for the garbage collector to find live
+1.650.230.6500 objects.

Stop-the-World Indicates that the garbage collector stops

www.azul.com
application processing to collect the heap.
[email protected]
Sweeping The garbage collection phase that locates
dead objects.
10