When it comes to Java applications, to make sure they run at peak performance, it’s critical to close the resource gap between the code and the virtual machine it’s running on – if there is one. The way to do that is by peaking into and fine-tuning the Java Virtual Machine (JVM).
However, that’s easier said than done. Many factors can influence JVM performance, and during tuning, you must consider all of them. Though, there are ways around that to make it not be such a pain.
In this post, we are going to help you plan and understand how to do JVM tuning by using the JVM parameters and following some performance tuning principles and procedures. Let’s get to it!
What Is JVM Tuning
Before diving into JVM tuning, there are a few things you should consider first:
- Cost – considering your environments, you may gain more by adding hardware instead of spending time to tune the JVM.
- Desired outcome – in the long run, it’s more efficient to tune for stability, rather than performance, but they do overlap.
- Ongoing issues – before starting tuning, you should do a thorough check of your system to see if there isn’t an underlying issue somewhere as tuning can delay or hide it temporarily. You can’t do JVM tuning, nor debugging if you’re not monitoring the JVM. Check out what are the critical JVM performance metrics to monitor and which are the best monitoring tools available today.
- Memory leaks – they always lead to garbage collection (GC) issues regardless of tuning.
Java virtual machine tuning is the process of adjusting the default parameters to match our application needs. This includes simple adjustments like the size of the heap, through choosing the right garbage collector to using optimized versions of getters.
App Optimization Considers All Performance Layers
Although critical, tuning JVM is not enough to ensure optimal performance. For instance, if an app has a badly designed architecture or poorly written code, you can’t expect performance to skyrocket by merely tuning the JVM.
Not to mention, once you have the first two optimized, you’d need to get back to the JVM and start it all over again. A well-done tuning looks at the whole system and all the layers that can impact performance, including database and OS.
That said, when you’re at the stage of performing JVM tuning, assume that the architecture and code of your project are either optimal or have been tuned. However, before diving into it, you must set performance optimization goals and identify the current performance issues. The goals will serve as a benchmark to compare the app with once it’s optimized and to determine if it needs further intervention.
Sematext JVM Monitoring covers all bases by finding bottlenecks in yours and 3rs party applications and services in real time.
Start your 30-day free trial See our plans
No credit card required – Get started in seconds!
What to Know Before Planning JVM Performance Tuning
Before diving into the step-by-step process, you need to get familiar with some basic concepts that we’ll use further on:
JVM Main Parameters
JVM parameters or JVM arguments are Java-specific values that change the behavior of the Java Virtual Machine.
When it comes to JVM tuning, there are few important parameters you should know as you’ll encounter them when configuring, tuning, and improving overall performance of the JVM.
Whatever you do, when it comes to JVM performance, you will, more than likely, reach a point where you have to initialize the heap memory. The parameters to specify minimum and maximum heap size are -Xms<heap size>[unit] and -Xmx<heap size>[unit], respectively. The unit is the unit you want to initialize the memory in. It can be ‘g’ (GB), ‘m’ (MB), or ‘k’ (KB).
When setting the minimum and maximum heap size of the JVM memory you may want to consider setting them to the same values. That way your heap will not have to be resized saving precious CPU cycles. You may also want to pre-touch all the pages if you are using larger heap by setting the -XX:+AlwaysPreTouch flag to the startup options.
Starting from Java 8 the Metaspace has replaced the old PermGen memory space. No more java.lang.OutOfMemoryError: PermGen errors, right now we can start monitoring our application logs for java.lang.OutOfMemoryError: Metadata space. By default the class metadata allocation is limited by the amount of available native memory and we have a new option to limit the maximum size of the new memory space introduced with Java 8. But just because the space can grow up to the limits of the native memory it doesn’t mean it is always occupying as much as it can. This region of the JVM memory can be dynamically resized depending on the needs.
The following properties exposed by the JVM allow us to control the Metaspace:
- -XX:MaxMetaspaceSize – unlimited by default, sets the maximum amount of native memory that can be allocated for the class metadata.
- -XX:MetaspaceSize – sets the size of the allocated class metadata above which the first garbage collection will be triggered. Default value is platform dependent.
- -XX:MinMetaspaceFreeRatio – minimum percentage of the Metaspace memory region that needs to be available after garbage collection. If the amount of memory left is below the threshold the Metaspace region will be resized.
- -XX:MaxMetaspaceFreeRatio – maximum percentage of the Metaspace memory region that needs to be available after the garbage collection. If the amount of memory left is above the threshold the Metaspace region will be resized.
Keep in mind that monitoring Metaspace region of the memory may be a very good idea as high garbage collection work in this memory region may indicate memory leaks in classes or classloaders.
Delving deeper, it’s important to take note of memory allocation parameters as well as you’ll need them during tuning. They are as follows:
- -XX:NewSize sets the initial size of young generation space
- -XXMaxNewSize states the maximum size of young generation space
- -Xmn specifies the size of the entire young generation space, meaning eden and the two survivor spaces
You will use these parameters to calculate aspects regarding the old generation space size:
- The size of the old generation is automatically set based on the size of the young generation space
- The initial old generation space equals the -Xmx value minus the -XX:NewSize value
- The minimum size of the old generation space equals the -Xmx value minus the -XX:MaxNewSize value
Out of Memory Error
OutOfMemoryError is probably every DevOps’ worst nightmare. You find yourself facing an app crash that is difficult to replicate and diagnose. Unfortunately, it’s quite frequent with large applications. Fortunately, JVM has parameters that write heap memory on a file that you can later use to troubleshoot.
-XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=./java_pid<pid>.hprof -XX:OnOutOfMemoryError="<cmd args>;<cmd args>" -XX:+UseGCOverheadLimit
- HeapDumpOnOutOfMemoryError commands the JVM to dump heap toa physical file when java.lang.OutOfMemoryError is thrown
- HeapDumpPath specifies the path to directory or filename
- OnOutOfMemoryError is used to run emergency user-defined commands when an OutOfMemoryError first occurs
- UseGCOverheadLimit is a policy that limits the proportion of the VM’s time that is spent in GC before an OutOfMemoryError is thrown
JVM has four garbage collector implementations:
- Serial Garbage Collector -XX:+UseSerialGC
- Parallel Garbage Collector -XX:+UseParallelGC
- CMS Garbage Collector -XX:+UseConcMarkSweepGC
- G1 Garbage Collector -XX:+UseG1GC
The first two garbage collectors are dedicated to work in the area of young generation heap space, while the CMS and G1 garbage collectors are dedicated for the old generation space.
Find out more about what Garbage Collection is and how it works (coming soon) and learn how to tune the garbage collector to improve app performance (coming soon).
Garbage Collection performance is closely related to JVM and app performance. When the garbage collector is not able to clear the memory it works more and more eventually leading to stop the world events or even out of memory situations. We want to avoid such situations as much as possible. To be able to do that we need to be able to observe what JVM garbage collector is doing. One of the best ways to monitor GC performance is to look at GC logs. You can log GC activity with the following commands:
-XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles=<number of log files> -XX:GCLogFileSize=<file size>[unit] -Xloggc:/path/to/gc.log
- -XX:+UseGCLogFileRotation specifies the log file rotation policy
- -XX:NumberOfGCLogFiles states the maximum number of log files to use when rotating logs
- -XX:GCLogFileSize refers to the maximum size of the log file
- -Xloggc specifies where the file is located
There are also additional JVM parameters important for GC logging. For example:
- -XX:+PrintGCDetails – includes additional details in the garbage collector log
- -XX:+PrintGCDateStamps – prints date stamps to the log
- -XX:+PrintGCTimeStamps – prints the timestamps to the log
- -XX:+PrintTenuringDistribution – adds information about tenuring age information to the log
- -XX:+PrintGCApplicationStoppedTime – includes information about how much time the application was stopped at safepoint which are commonly caused by stop the world garbage collection
- -XX:+UseStringDeduplication eliminates duplicate string value to a single instance to reduce unnecessary memory usage.
- -XX:+UseLWPSynchronization enables LWP (Light Weight Process) – based instead of thread-based synchronization.
- -XX:LargePageSizeInBytes sets the large page size used for the Java heap. Have in mind, though, that while larger page sizes allow for more efficient use of memory hardware, this may lead to larger space sizes for Metaspace, which can force you to reduce the size of the Java heap space.
- -XX:MaxHeapFreeRatio enables the maximum percentage of heap free after GC to avoid shrinking.
- -XX:MinHeapFreeRatio enables the minimum percentage of heap free after GC to avoid expansion.
- -XX:SurvivorRatio sets the ratio of eden/survivor space size.
- -XX:+UseLargePages use large page memory.
- -XX:+UseStringCache enables caching of commonly allocated strings.
- -XX:+UseCompressedStrings use a byte type for String objects, which can be represented in pure ASCII format.
- -XX:+OptimizeStringConcat optimizes String concatenation operations where possible.
With Sematext JVM Monitoring, you get detailed visibility of heap sizes and utilization for individual memory pool to help spot performance issues faster.
Start your 30-day free trial See our plans
No credit card required – Get started in seconds!
Set your JVM Performance Goals
Before you start tuning the JVM for performance, you first have to set performance goals. You will set values for these goals later on:
- Latency is the amount of time required to run a garbage collection event.
- Throughput is the percentage of time the VM spends executing the application versus time spent performing garbage collection.
- Footprint is the amount of memory required by the garbage collector to run smoothly.
However, you can’t focus on all three goals at once, as any performance gain from any of them results in a performance loss of the other one or two. For example:
- High throughput and low latency results in higher memory usage.
- High throughput and low memory usage results in higher latency.
- Low latency and low memory usage results in lower throughput.
You have to decide which two are the most relevant to your app while you consider your business needs. Either way, the goal of JVM tuning is to optimize the garbage collector so that you have high throughput with less memory consumption and low latency. However, less memory/low latency does not automatically mean that the less or lower the memory or latency is, the better the performance. That depends on which metrics you choose to keep an eye on.
JVM Tuning Principles
When performing performance tuning, have in mind the following principles as they make garbage collection easier.
- Minor GC collection – implies that Minor GC should collect as many dead objects as possible to reduce the frequency of Full GC.
- GC memory maximization – it says that the more memory the GC can access during a cycle, the more efficient the cleanup and the lower the collection frequency.
- Two out of three – as stated above, you need to pick two out of the three performance goals.
JVM Tuning Procedure
First, what you need to have in mind is that Java VM tuning can’t solve all performance issues. Therefore, it should be done only necessary. That said, tuning is a lengthy process where you’ll more than likely perform ongoing configuration optimizations and multiple iterations based on the stress and benchmark tests results. Before meeting the desired metric, you may also need to tune parameters many times, thus running the tests all over again.
As a general rule, tuning should first meet the memory usage requirements, latency, and finally, throughput, as you’ll see in the following sections.
We are going to use the default parallel garbage collector (or throughput collector) in JDK, which you can enable with -XX:+UseParallelGC. This flag enables a parallel version of both the young- and old-generation collectors. You can check the defaults for your Java version by running java -XX:+PrintCommandLineFlags -version. You can also the serial garbage collector by using the -XX:+UseSerialGC. The serial collector is a single-thread collector, while the parallel collector is a multiple-thread collector.
How to Do JVM Performance Tuning
Once you know all of the above, here are the steps you need to take to perform JVM tuning:
Measure Memory Footprint
To determine the memory usage, you first need to know the size of the active data. The size of the active data is the amount of Java heap that is occupied by data that was active since the application entered the steady stage.
Application Phase: Stable vs. Start-up
Active data has to be measured in a stable state instead of the start-up phase. In the start-up phase, the JVM loads and launches the app’s main modules and data; thus, the JVM parameters are not yet stable.
On the other hand, the stable phase implies that an app has already been running for some time and underwent a stress test. More specifically, an app is in a stable phase when it reaches a workload that meets the requirements at the business peak in the production environment and stays steady after the peak is reached. Only then every JVM performance parameter is in a stable state.
How to Determine Memory Footprint
Make sure to perform the tests with the default JVM parameters as it allows you to see how much memory the app requires while in the stable phase. The Full GC
Once the app is running in a stable state, you have to estimate the memory footprint based on the average old and permanent occupancy, you can see with Full GC logs during stable state periods. You can also use the longest Full GC for estimation.
GC logs are one of the best ways to gather meaningful and rich data to help with tuning. Enabling GC logs doesn’t affect performance. Therefore you can use them even in production environments to detect issues.
Use the following command-line to turn on Full GC logs:
-XX:+PrintGC -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -Xloggc:<filename>
However, if there are no Full GC logs available, you can use monitoring tools to call for them or the following command to enable them:
jmap -histo:live pid
Either way, you get information similar to:
0.134: [GC (Allocation Failure) [PSYoungGen: 65536K->10720K(76288K)] 65536K->40488K(251392K), 0.0190287 secs] [Times: user=0.13 sys=0.04, real=0.02 secs] 0.193: [GC (Allocation Failure) [PSYoungGen: 71912K->10752K(141824K)] 101680K->101012K(316928K), 0.0357512 secs] [Times: user=0.27 sys=0.06, real=0.04 secs] 0.374: [GC (Allocation Failure) [PSYoungGen: 141824K->10752K(141824K)] 232084K->224396K(359424K), 0.0809666 secs] [Times: user=0.58 sys=0.12, real=0.08 secs] 0.455: [Full GC (Ergonomics) [PSYoungGen: 10752K->0K(141824K)] [ParOldGen: 213644K->215361K(459264K)] 224396K->215361K(601088K), [Metaspace: 2649K->2649K(1056768K)], 0.4409247 secs] [Times: user=3.46 sys=0.02, real=0.44 secs] 0.984: [GC (Allocation Failure) [PSYoungGen: 131072K->10752K(190464K)] 346433K->321225K(649728K), 0.1407158 secs] [Times: user=1.28 sys=0.08, real=0.14 secs] 1.168: [GC (System.gc()) [PSYoungGen: 60423K->10752K(190464K)] 370896K->368961K(649728K), 0.0676498 secs] [Times: user=0.53 sys=0.05, real=0.06 secs] 1.235: [Full GC (System.gc()) [PSYoungGen: 10752K->0K(190464K)] [ParOldGen: 358209K->368152K(459264K)] 368961K->368152K(649728K), [Metaspace: 2652K->2652K(1056768K)], 1.1751101 secs] [Times: user=10.64 sys=0.05, real=1.18 secs] 2.612: [Full GC (Ergonomics) [PSYoungGen: 179712K->0K(190464K)] [ParOldGen: 368152K->166769K(477184K)] 547864K->166769K(667648K), [Metaspace: 2659K->2659K(1056768K)], 0.2662589 secs] [Times: user=2.14 sys=0.00, real=0.27 secs]
We can see several operations performed by our JVM version 8. Let’s look at one of the lines here that described the Full GC event performed on purpose from our test code by using the System.gc() method:
1.235: [Full GC (System.gc()) [PSYoungGen: 10752K->0K(190464K)] [ParOldGen: 358209K->368152K(459264K)] 368961K->368152K(649728K), [Metaspace: 2652K->2652K(1056768K)], 1.1751101 secs] [Times: user=10.64 sys=0.05, real=1.18 secs]
As you can see there is a lot of information in a single line describing the Full GC event. Apart from the event type we see what was happening in the young generation space [PSYoungGen: 10752K->0K(190464K)], what was happening in the old generation space, [ParOldGen: 358209K->368152K(459264K)], what was happening in the Metaspace region of memory [Metaspace: 2652K->2652K(1056768K)], the total memory difference 368961K->368152K(649728K) and finally the time the whole operation took: [Times: user=10.64 sys=0.05, real=1.18 secs].
The information about young generation garbage collector work tells us that after the young generation garbage collection event the space went down from 10752K to 0K and the total allocated young generation space is 190464K.
The old generation garbage collector started with the memory usage at 358209K and after its work finished with 368152K. The total memory allocated to old generation space is 459264K. This means that this GC cycle didn’t end up freeing too much old generation space.
After the whole garbage collection was done we ended up with 368152K memory starting from the initial 368961K and the whole memory space occupied was 649728K.
The Metaspace region of the memory started and ended with the same memory usage 2652K and the whole region occupies 1056768K.
Finally, it took JVM 1.18 seconds to perform the whole garbage collection operation. The user=10.64 section tells us the amount of CPU time spent in user-mode code outside of the operating system kernel. The sys=0.05 section is the CPU time spent inside the kernel within the process itself, which means how long the CPU spent executing system related calls.
As you can see a single garbage collector log line can give us a lot of insight into what happened, how much memory was released and how long the whole operation took.
Tuning for Latency
Once the memory footprint established, the next step is latency tuning. At this stage, heap memory size and latency don’t meet application requirements. Therefore, a new debugging is needed with the actual requirements that the app has. You might have to tune the heap size again, determine GC duration and frequency, and decide if you need to switch to another garbage collector.
Determine System Latency Requirements
Previously, we’ve mentioned performance goals, but we didn’t set values for them. These goals serve as system latency requirements you need to meet after tuning. The metrics that help get there are:
- Acceptable average Minor GC frequency, which you’ll compare with the number of Minor GCs.
- Acceptable maximum Full GC pause, which you’ll compare with the longest Full GC cycle.
- Acceptable frequency of maximum Full GC pause, which you’ll compare with the highest frequency of the Full GC.
- Acceptable average Minor GC pause, which you’ll compare with the Minor GC duration.
You can get this data by optimizing the young generation and old generation size, respectively. You can read more about it in our guide to garbage collection tuning.
Tuning for Throughput
In the final step of JVM performance tuning, we run throughput tests on the results we got so far and then make slight adjustments if needed.
Based on tests and overall application requirements, an app should have a set throughput metric. When this goal is met or exceeded even, you can stop the tuning.
However, if after optimization, you still can’t reach the throughput goal, you need to get back to it and assess the gap between the throughput requirement and the current throughout. If the gap is around 20%, you can change the parameters, increase memory, and debug the application again. However, if the gap is larger than 20%, you need to review the throughput goal as the throughput goal, and the design may not meet the requirements of the whole Java application.
For a garbage collection, throughput tuning has two purposes: to minimize the number of objects passed to the old area and decrease Full GC execution time or Stop-the-World events. These can lead to low throughput.
Sematext JVM Monitoring makes monitoring easy by giving you GC, thread, and memory usage monitoring out-of-the-box.
Start your 30-day free trial See our plans
No credit card required – Get started in seconds!
Time to Tune!
JVM performance tuning is a complex process that requires a methodical approach and a clear idea of the desired performance outcomes, as you’ve noticed in this JVM tuning guide. The default JVM tunings do a great job of improving application performance and making sure it’s stable in most situations. On the other hand, depending on how your app handles the required workload, the basic initial JVM parameters may not be enough. In that case, you’ll have to benchmark your app and tune the JVM to reach an appropriate set of parameters.