Do not choose a maximum value for the heap unless you know that you need a heap greater than the default maximum heap size. Choose a throughput goal that is sufficient for your application.
The heap will grow or shrink to a size that will support the chosen throughput goal. A change in the application's behavior can cause the heap to grow or shrink. For example, if the application starts allocating at a higher rate, the heap will grow to maintain the same throughput.
If the heap grows to its maximum size and the throughput goal is not being met, the maximum heap size is too small for the throughput goal. Set the maximum heap size to a value that is close to the total physical memory on the platform but which does not cause swapping of the application. Execute the application again. If the throughput goal is still not met, then the goal for the application time is too high for the available memory on the platform.
If the throughput goal can be met, but there are pauses that are too long, then select a maximum pause time goal. Choosing a maximum pause time goal may mean that your throughput goal will not be met, so choose values that are an acceptable compromise for the application.
It is typical that the size of the heap will oscillate as the garbage collector tries to satisfy competing goals. This is true even if the application has reached a steady state. The pressure to achieve a throughput goal (which may require a larger heap) competes with the goals for a maximum pause time and a minimum footprint (which both may require a small heap).
Page 2
One strength of the Java SE platform is that it shields the developer from the complexity of memory allocation and garbage collection. However, when garbage collection is the principal bottleneck, it is useful to understand some aspects of this hidden implementation. Garbage collectors make assumptions about the way applications use objects, and these are reflected in tunable parameters that can be adjusted for improved performance without sacrificing the power of the abstraction.
An object is considered garbage when it can no longer be reached from any pointer in the running program. The most straightforward garbage collection algorithms iterate over every reachable object. Any objects left over are considered garbage. The time this approach takes is proportional to the number of live objects, which is prohibitive for large applications maintaining lots of live data.
The virtual machine incorporates a number of different garbage collection algorithms that are combined using generational collection. While naive garbage collection examines every live object in the heap, generational collection exploits several empirically observed properties of most applications to minimize the work required to reclaim unused (garbage) objects. The most important of these observed properties is the weak generational hypothesis, which states that most objects survive for only a short period of time.
The blue area in Figure 3-1, "Typical Distribution for Lifetimes of Objects" is a typical distribution for the lifetimes of objects. The x-axis is object lifetimes measured in bytes allocated. The byte count on the y-axis is the total bytes in objects with the corresponding lifetime. The sharp peak at the left represents objects that can be reclaimed (in other words, have "died") shortly after being allocated. Iterator objects, for example, are often alive for the duration of a single loop.
Some objects do live longer, and so the distribution stretches out to the right. For instance, there are typically some objects allocated at initialization that live until the process exits. Between these two extremes are objects that live for the duration of some intermediate computation, seen here as the lump to the right of the initial peak. Some applications have very different looking distributions, but a surprisingly large number possess this general shape. Efficient collection is made possible by focusing on the fact that a majority of objects "die young."
To optimize for this scenario, memory is managed in generations (memory pools holding objects of different ages). Garbage collection occurs in each generation when the generation fills up. The vast majority of objects are allocated in a pool dedicated to young objects (the young generation), and most objects die there. When the young generation fills up, it causes a minor collection in which only the young generation is collected; garbage in other generations is not reclaimed. Minor collections can be optimized, assuming that the weak generational hypothesis holds and most objects in the young generation are garbage and can be reclaimed. The costs of such collections are, to the first order, proportional to the number of live objects being collected; a young generation full of dead objects is collected very quickly. Typically, some fraction of the surviving objects from the young generation are moved to the tenured generation during each minor collection. Eventually, the tenured generation will fill up and must be collected, resulting in a major collection, in which the entire heap is collected. Major collections usually last much longer than minor collections because a significantly larger number of objects are involved.
As noted in the section Ergonomics, ergonomics selects the garbage collector dynamically to provide good performance on a variety of applications. The serial garbage collector is designed for applications with small data sets, and its default parameters were chosen to be effective for most small applications. The parallel or throughput garbage collector is meant to be used with applications that have medium to large data sets. The heap size parameters selected by ergonomics plus the features of the adaptive size policy are meant to provide good performance for server applications. These choices work well in most, but not all, cases, which leads to the central tenet of this document:
Note:
If garbage collection becomes a bottleneck, you will most likely have to customize the total heap size as well as the sizes of the individual generations. Check the verbose garbage collector output and then explore the sensitivity of your individual performance metric to the garbage collector parameters.
Figure 3-2, "Default Arrangement of Generations, Except for Parallel Collector and G1" shows the default arrangement of generations (for all collectors with the exception of the parallel collector and G1):
At initialization, a maximum address space is virtually reserved but not allocated to physical memory unless it is needed. The complete address space reserved for object memory can be divided into the young and tenured generations.
The young generation consists of eden and two survivor spaces. Most objects are initially allocated in eden. One survivor space is empty at any time, and serves as the destination of any live objects in eden; the other survivor space is the destination during the next copying collection. Objects are copied between survivor spaces in this way until they are old enough to be tenured (copied to the tenured generation).
There are two primary measures of garbage collection performance:
- Throughput is the percentage of total time not spent in garbage collection considered over long periods of time. Throughput includes time spent in allocation (but tuning for speed of allocation is generally not needed).
- Pauses are the times when an application appears unresponsive because garbage collection is occurring.
Users have different requirements of garbage collection. For example, some consider the right metric for a web server to be throughput because pauses during garbage collection may be tolerable or simply obscured by network latencies. However, in an interactive graphics program, even short pauses may negatively affect the user experience.
Some users are sensitive to other considerations. Footprint is the working set of a process, measured in pages and cache lines. On systems with limited physical memory or many processes, footprint may dictate scalability. Promptness is the time between when an object becomes dead and when the memory becomes available, an important consideration for distributed systems, including Remote Method Invocation (RMI).
In general, choosing the size for a particular generation is a trade-off between these considerations. For example, a very large young generation may maximize throughput, but does so at the expense of footprint, promptness, and pause times. Young generation pauses can be minimized by using a small young generation at the expense of throughput. The sizing of one generation does not affect the collection frequency and pause times for another generation.
There is no one right way to choose the size of a generation. The best choice is determined by the way the application uses memory as well as user requirements. Thus the virtual machine's choice of a garbage collector is not always optimal and may be overridden with command-line options described in the section Sizing the Generations.
Throughput and footprint are best measured using metrics particular to the application. For example, the throughput of a web server may be tested using a client load generator, whereas the footprint of the server may be measured on the Solaris operating system using the pmap command. However, pauses due to garbage collection are easily estimated by inspecting the diagnostic output of the virtual machine itself.
The command-line option -verbose:gc causes information about the heap and garbage collection to be printed at each collection. For example, here is output from a large server application:
[GC 325407K->83000K(776768K), 0.2300771 secs] [GC 325816K->83372K(776768K), 0.2454258 secs] [Full GC 267628K->83769K(776768K), 1.8479984 secs]
The output shows two minor collections followed by one major collection. The numbers before and after the arrow (for example, 325407K->83000K from the first line) indicate the combined size of live objects before and after garbage collection, respectively. After minor collections, the size includes some objects that are garbage (no longer alive) but cannot be reclaimed. These objects are either contained in the tenured generation or referenced from the tenured generation.
The next number in parentheses (for example, (776768K) again from the first line) is the committed size of the heap: the amount of space usable for Java objects without requesting more memory from the operating system. Note that this number only includes one of the survivor spaces. Except during a garbage collection, only one survivor space will be used at any given time to store objects.
The last item on the line (for example, 0.2300771 secs) indicates the time taken to perform the collection, which is in this case approximately a quarter of a second.
The format for the major collection in the third line is similar.
Note:
The format of the output produced by -verbose:gc is subject to change in future releases.
The command-line option -XX:+PrintGCDetails causes additional information about the collections to be printed. An example of the output with -XX:+PrintGCDetails using the serial garbage collector is shown here.
[GC [DefNew: 64575K->959K(64576K), 0.0457646 secs] 196016K->133633K(261184K), 0.0459067 secs]
This indicates that the minor collection recovered about 98% of the young generation, DefNew: 64575K->959K(64576K) and took 0.0457646 secs (about 45 milliseconds).
The usage of the entire heap was reduced to about 51% (196016K->133633K(261184K)), and there was some slight additional overhead for the collection (over and above the collection of the young generation) as indicated by the final time of 0.0459067 secs.
Note:
The format of the output produced by -XX:+PrintGCDetails is subject to change in future releases.
The option -XX:+PrintGCTimeStamps adds a time stamp at the start of each collection. This is useful to see how frequently garbage collections occur.
111.042: [GC 111.042: [DefNew: 8128K->8128K(8128K), 0.0000505 secs]111.042: [Tenured: 18154K->2311K(24576K), 0.1290354 secs] 26282K->2311K(32704K), 0.1293306 secs]
The collection starts about 111 seconds into the execution of the application. The minor collection starts at about the same time. Additionally, the information is shown for a major collection delineated by Tenured. The tenured generation usage was reduced to about 10% (18154K->2311K(24576K)) and took 0.1290354 secs (approximately 130 milliseconds).
Page 3
The following discussion regarding growing and shrinking of the heap and default heap sizes does not apply to the parallel collector. (See the section Parallel Collector Ergonomics in Sizing the Generations for details on heap resizing and default heap sizes with the parallel collector.) However, the parameters that control the total size of the heap and the sizes of the generations do apply to the parallel collector.
The most important factor affecting garbage collection performance is total available memory. Because collections occur when generations fill up, throughput is inversely proportional to the amount of memory available.
By default, the virtual machine grows or shrinks the heap at each collection to try to keep the proportion of free space to live objects at each collection within a specific range. This target range is set as a percentage by the parameters -XX:MinHeapFreeRatio=``<minimum> and -XX:MaxHeapFreeRatio=``<maximum>, and the total size is bounded below by -Xms``<min> and above by -Xmx``<max>. The default parameters for the 64-bit Solaris operating system (SPARC Platform Edition) are shown in Table 4-1, "Default Parameters for 64-Bit Solaris Operating System":
With these parameters, if the percent of free space in a generation falls below 40%, then the generation will be expanded to maintain 40% free space, up to the maximum allowed size of the generation. Similarly, if the free space exceeds 70%, then the generation will be contracted so that only 70% of the space is free, subject to the minimum size of the generation.
As noted in Table 4-1, "Default Parameters for 64-Bit Solaris Operating System", the default maximum heap size is a value that is calculated by the JVM. The calculation used in Java SE for the parallel collector and the server JVM are now used for all the garbage collectors. Part of the calculation is an upper limit on the maximum heap size that is different for 32-bit platforms and 64-bit platforms. See the section Default Heap Size in The Parallel Collector. There is a similar calculation for the client JVM, which results in smaller maximum heap sizes than for the server JVM.
The following are general guidelines regarding heap sizes for server applications:
- Unless you have problems with pauses, try granting as much memory as possible to the virtual machine. The default size is often too small.
- Setting
-Xmsand-Xmxto the same value increases predictability by removing the most important sizing decision from the virtual machine. However, the virtual machine is then unable to compensate if you make a poor choice.
- In general, increase the memory as you increase the number of processors, since allocation can be parallelized.
Page 4
Unless your application has rather strict pause time requirements, first run your application and allow the VM to select a collector. If necessary, adjust the heap size to improve performance. If the performance still does not meet your goals, then use the following guidelines as a starting point for selecting a collector.
- If the application has a small data set (up to approximately 100 MB), then
select the serial collector with the option `-XX:+UseSerialGC`.
- If the application will be run on a single processor and there are no pause time requirements, then let the VM select the collector, or select the serial collector with the option
-XX:+UseSerialGC.
- If (a) peak application performance is the first priority and (b) there are no pause time requirements or pauses of 1 second or longer are acceptable, then let the VM select the collector, or select the parallel collector with
-XX:+UseParallelGC.
- If response time is more important than overall throughput and garbage collection pauses must be kept shorter than approximately 1 second, then select the concurrent collector with
-XX:+UseConcMarkSweepGCor-XX:+UseG1GC.
These guidelines provide only a starting point for selecting a collector because performance is dependent on the size of the heap, the amount of live data maintained by the application, and the number and speed of available processors. Pause times are particularly sensitive to these factors, so the threshold of 1 second mentioned previously is only approximate: the parallel collector will experience pause times longer than 1 second on many data size and hardware combinations; conversely, the concurrent collector may not be able to keep pauses shorter than 1 second on some combinations.
If the recommended collector does not achieve the desired performance, first attempt to adjust the heap and generation sizes to meet the desired goals. If performance is still inadequate, then try a different collector: use the concurrent collector to reduce pause times and use the parallel collector to increase overall throughput on multiprocessor hardware.
Page 5
The statistics such as average pause time kept by the collector are updated at the end of each collection. The tests to determine if the goals have been met are then made and any needed adjustments to the size of a generation is made. The exception is that explicit garbage collections (for example, calls to System.gc()) are ignored in terms of keeping statistics and making adjustments to the sizes of generations.
Growing and shrinking the size of a generation is done by increments that are a fixed percentage of the size of the generation so that a generation steps up or down toward its desired size. Growing and shrinking are done at different rates. By default a generation grows in increments of 20% and shrinks in increments of 5%. The percentage for growing is controlled by the command-line option -XX:YoungGenerationSizeIncrement=``<Y> for the young generation and -XX:TenuredGenerationSizeIncrement=``<T> for the tenured generation. The percentage by which a generation shrinks is adjusted by the command-line flag -XX:AdaptiveSizeDecrementScaleFactor=``<D>. If the growth increment is X percent, then the decrement for shrinking is X/D percent.
If the collector decides to grow a generation at startup, then there is a supplemental percentage is added to the increment. This supplement decays with the number of collections and has no long-term effect. The intent of the supplement is to increase startup performance. There is no supplement to the percentage for shrinking.
If the maximum pause time goal is not being met, then the size of only one generation is shrunk at a time. If the pause times of both generations are above the goal, then the size of the generation with the larger pause time is shrunk first.
If the throughput goal is not being met, the sizes of both generations are increased. Each is increased in proportion to its respective contribution to the total garbage collection time. For example, if the garbage collection time of the young generation is 25% of the total collection time and if a full increment of the young generation would be by 20%, then the young generation would be increased by 5%.
Page 6
The mostly concurrent collector trades processor resources (which would otherwise be available to the application) for shorter major collection pause times. The most visible overhead is the use of one or more processors during the concurrent parts of the collection. On an N processor system, the concurrent part of the collection will use K/N of the available processors, where 1<=K<=ceiling{N/4}. (Note that the precise choice of and bounds on K are subject to change.) In addition to the use of processors during concurrent phases, additional overhead is incurred to enable concurrency. Thus while garbage collection pauses are typically much shorter with the concurrent collector, application throughput also tends to be slightly lower than with the other collectors.
On a machine with more than one processing core, processors are available for application threads during the concurrent part of the collection, so the concurrent garbage collector thread does not "pause" the application. This usually results in shorter pauses, but again fewer processor resources are available to the application and some slowdown should be expected, especially if the application uses all of the processing cores maximally. As N increases, the reduction in processor resources due to concurrent garbage collection becomes smaller, and the benefit from concurrent collection increases. The section Concurrent Mode Failure in Concurrent Mark Sweep (CMS) Collector discusses potential limits to such scaling.
Because at least one processor is used for garbage collection during the concurrent phases, the concurrent collectors do not normally provide any benefit on a uniprocessor (single-core) machine. However, there is a separate mode available for CMS (not G1) that can achieve low pauses on systems with only one or two processors; see Incremental Mode in Concurrent Mark Sweep (CMS) Collector for details. This feature is being deprecated in Java SE 8 and may be removed in a later major release.
Page 7
Note that the incremental mode is being deprecated in Java SE 8 and may be removed in a future major release.
The CMS collector can be used in a mode in which the concurrent phases are done incrementally. Recall that during a concurrent phase the garbage collector thread is using one or more processors. The incremental mode is meant to lessen the effect of long concurrent phases by periodically stopping the concurrent phase to yield back the processor to the application. This mode, referred to here as i-cms, divides the work done concurrently by the collector into small chunks of time that are scheduled between young generation collections. This feature is useful when applications that need the low pause times provided by the CMS collector are run on machines with small numbers of processors (for example, 1 or 2).
The concurrent collection cycle typically includes the following steps:
- Stop all application threads, identify the set of objects reachable from roots, and then resume all application threads.
- Concurrently trace the reachable object graph, using one or more processors, while the application threads are executing.
- Concurrently retrace sections of the object graph that were modified since the tracing in the previous step, using one processor.
- Stop all application threads and retrace sections of the roots and object graph that may have been modified since they were last examined, and then resume all application threads.
- Concurrently sweep up the unreachable objects to the free lists used for allocation, using one processor.
- Concurrently resize the heap and prepare the support data structures for the next collection cycle, using one processor.
Normally, the CMS collector uses one or more processors during the entire concurrent tracing phase, without voluntarily relinquishing them. Similarly, one processor is used for the entire concurrent sweep phase, again without relinquishing it. This overhead can be too much of a disruption for applications with response time constraints that might otherwise have used the processing cores, particularly when run on systems with just one or two processors. Incremental mode solves this problem by breaking up the concurrent phases into short bursts of activity, which are scheduled to occur midway between minor pauses.
The i-cms mode uses a duty cycle to control the amount of work the CMS collector is allowed to do before voluntarily giving up the processor. The duty cycle is the percentage of time between young generation collections that the CMS collector is allowed to run. The i-cms mode can automatically compute the duty cycle based on the behavior of the application (the recommended method, known as automatic pacing), or the duty cycle can be set to a fixed value on the command line.
To use i-cms in Java SE 8, use the following command-line options:
-XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode \ -XX:+PrintGCDetails -XX:+PrintGCTimeStamps
The first two options enable the CMS collector and i-cms, respectively. The last two options are not required; they simply cause diagnostic information about garbage collection to be written to standard output, so that garbage collection behavior can be seen and later analyzed.
For Java SE 5 and earlier releases, Oracle recommends using the following as an initial set of command-line options for i-cms:
-XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode \ -XX:+PrintGCDetails -XX:+PrintGCTimeStamps \ -XX:+CMSIncrementalPacing -XX:CMSIncrementalDutyCycleMin=0 -XX:CMSIncrementalDutyCycle=10
The same values are recommended for JavaSE8 although the values for the three options that control i-cms automatic pacing became the default in JavaSE6.
Page 8
The Garbage-First (G1) garbage collector is a server-style garbage collector, targeted for multiprocessor machines with large memories. It attempts to meet garbage collection (GC) pause time goals with high probability while achieving high throughput. Whole-heap operations, such as global marking, are performed concurrently with the application threads. This prevents interruptions proportional to heap or live-data size.
The G1 collector achieves high performance and pause time goals through several techniques.
The heap is partitioned into a set of equally sized heap regions, each a contiguous range of virtual memory. G1 performs a concurrent global marking phase to determine the liveness of objects throughout the heap. After the marking phase completes, G1 knows which regions are mostly empty. It collects these regions first, which often yields a large amount of free space. This is why this method of garbage collection is called Garbage-First. As the name suggests, G1 concentrates its collection and compaction activity on the areas of the heap that are likely to be full of reclaimable objects, that is, garbage. G1 uses a pause prediction model to meet a user-defined pause time target and selects the number of regions to collect based on the specified pause time target.
G1 copies objects from one or more regions of the heap to a single region on the heap, and in the process both compacts and frees up memory. This evacuation is performed in parallel on multiprocessors to decrease pause times and increase throughput. Thus, with each garbage collection, G1 continuously works to reduce fragmentation. This is beyond the capability of both of the previous methods. CMS (Concurrent Mark Sweep) garbage collection does not do compaction. Parallel compaction performs only whole-heap compaction, which results in considerable pause times.
It is important to note that G1 is not a real-time collector. It meets the set pause time target with high probability but not absolute certainty. Based on data from previous collections, G1 estimates how many regions can be collected within the target time. Thus, the collector has a reasonably accurate model of the cost of collecting the regions, and it uses this model to determine which and how many regions to collect while staying within the pause time target.
The first focus of G1 is to provide a solution for users running applications that require large heaps with limited GC latency. This means heap sizes of around 6 GB or larger, and a stable and predictable pause time below 0.5 seconds.
Applications running today with either the CMS or the with parallel compaction would benefit from switching to G1 if the application has one or more of the following traits.
- More than 50% of the Java heap is occupied with live data.
- The rate of object allocation rate or promotion varies significantly.
- The application is experiencing undesired long garbage collection or compaction pauses (longer than 0.5 to 1 second).
G1 is planned as the long-term replacement for the Concurrent Mark-Sweep Collector (CMS). Comparing G1 with CMS reveals differences that make G1 a better solution. One difference is that G1 is a compacting collector. Also, G1 offers more predictable garbage collection pauses than the CMS collector, and allows users to specify desired pause targets.
As with CMS, G1 is designed for applications that require shorter GC pauses.
G1 divides the heap into fixed-sized regions (the gray boxes) as in Figure 9-1, "Heap Division by G1".
G1 is generational in a logical sense. A set of empty regions is designated as the logical young generation. In the figure, the young generation is light blue. Allocations are done out of that logical young generation, and when the young generation is full, that set of regions is garbage collected (a young collection). In some cases, regions outside the set of young regions (old regions in dark blue) can be garbage collected at the same time. This is referred to as a mixed collection. In the figure, the regions being collected are marked by red boxes. The figure illustrates a mixed collection because both young regions and old regions are being collected. The garbage collection is a compacting collection that copies live objects to selected, initially empty regions. Based on the age of a surviving object, the object can be copied to a survivor region (marked by "S") or to an old region (not specifically shown). The regions marked by "H" contain humongous objects that are larger than half a region and are treated specially; see the section Humongous Objects and Humongous Allocations in Garbage-First Garbage Collector.
As with CMS, the G1 collector runs parts of its collection while the application continues to run and there is a risk that the application will allocate objects faster than the garbage collector can recover free space. See the section Concurrent Mode Failure in Concurrent Mark Sweep (CMS) Collector for the analogous CMS behavior. In G1, the failure (exhaustion of the Java heap) occurs while G1 is copying live data out of one region (evacuating) into another region. The copying is done to compact the live data. If a free (empty) region cannot be found during the evacuation of a region being garbage collected, then an allocation failure occurs (because there is no space to allocate the live objects from the region being evacuated) and a stop-the-world (STW) full collection is done.
G1 pauses the application to copy live objects to new regions. These pauses can either be young collection pauses where only young regions are collected or mixed collection pauses where young and old regions are evacuated. As with CMS there is a final marking or remark pause to complete the marking while the application is stopped. Whereas CMS also had an initial marking pause, G1 does the initial marking work as part of an evacuation pause. G1 has a cleanup phase at the end of a collection which is partly STW and partly concurrent. The STW part of the cleanup phase identifies empty regions and determines old regions that are candidates for the next collection.
Set a pause time goal for G1 with the flag MaxGCPauseMillis. G1 uses a prediction model to decide how much garbage collection work can be done within that target pause time. At the end of a collection, G1 chooses the regions to be collected in the next collection (the collection set). The collection set will contain young regions (the sum of whose sizes determines the size of the logical young generation). It is partly through the selection of the number of young regions in the collection set that G1 exerts control over the length of the GC pauses. You can specify the size of the young generation on the command line as with the other garbage collectors, but doing so may hamper the ability of G1 to attain the target pause time. In addition to the pause time goal, you can specify the length of the time period during which the pause can occur. You can specify the minimum mutator usage with this time span (GCPauseIntervalMillis) along with the pause time goal. The default value for MaxGCPauseMillis is 200 milliseconds. The default value for GCPauseIntervalMillis (0) is the equivalent of no requirement on the time span.