Other Examples

Previous chapters covered the most common problems related to poorly behaving GC. Unfortunately, there is a long list of more specific cases, where you cannot apply the knowledge from previous chapters. This section describes a few of the more unusual problems that you may face.

RMI & GC

When your application is publishing or consuming services over RMI, the JVM periodically launches full GC to make sure that locally unused objects are also not taking up space on the other end. Bear in mind that even if you are not explicitly publishing anything over RMI in your code, third party libraries or utilities can still open RMI endpoints. One such common culprit is for example JMX, which, if attached to remotely, will use RMI underneath to publish the data.

The problem is exposed by seemingly unnecessary and periodic full GC pauses. When you check the old generation consumption, there is often no pressure to the memory as there is plenty of free space in the old generation, but full GC is triggered, stopping the application threads.

This behavior of removing remote references via System.gc() is triggered by the sun.rmi.transport.ObjectTable class requesting garbage collection to be run periodically as specified in the sun.misc.GC.requestLatency() method.

For many applications, this is not necessary or outright harmful. To disable such periodic GC runs, you can set up the following for your JVM startup scripts:

java -Dsun.rmi.dgc.server.gcInterval=9223372036854775807L -Dsun.rmi.dgc.client.gcInterval=9223372036854775807L com.yourcompany.YourApplication

This sets the period after which System.gc() is run to Long.MAX_VALUE; for all practical matters, this equals eternity.

An alternative solution for the problem would to disable explicit calls to System.gc() by specifying -XX:+DisableExplicitGC in the JVM startup parameters. We do not however recommend this solution as it can have other side effects.

JVMTI tagging & GC

Whenever the application is run alongside with a Java Agent (-javaagent), there is a chance that the agent can tag the objects in the heap using JVMTI tagging. Agents can use tagging for various reasons that are not in the scope of this handbook, but there is a GC-related performance issue that can start affecting the latency and throughput of your application if tagging is applied to a large subset of objects inside the heap.

The problem is hidden in the native code where JvmtiTagMap::do_weak_oops iterates over all the tags during each garbage collection event and performs a number of not-so-cheap operations for all of them. To make things worse, this operation is performed sequentially and is not parallelized.

With a large number of tags, this implies that a large part of the GC process is now carried out in a single thread and all the benefits of parallelism disappear, potentially increasing the duration of GC pauses by an order of magnitude.

To check whether or not a particular agent can be the reason for extended GC pauses, you would need to turn on the diagnostic option of –XX:+TraceJVMTIObjectTagging. Enabling the trace will allow you to get an estimate of how much native memory the tag map consumes and how much time the heap walks take.

If you are not the author of the agent yourself, fixing the problem is often out of your reach. Apart from contacting the vendor of a particular agent you cannot do much. In case you do end up in a situation like this, recommend that the vendor clean up the tags that are no longer needed.

Humongous Allocations

Whenever your application is using the G1 garbage collection algorithm, a phenomenon called humongous allocations can impact your application performance in regards of GC. To recap, humongous allocations are allocations that are larger than 50% of the region size in G1.

Having frequent humongous allocations can trigger GC performance issues, considering the way that G1 handles such allocations:

  • If the regions contain humongous objects, space between the last humongous object in the region and the end of the region will be unused. If all the humongous objects are just a bit larger than a factor of the region size, this unused space can cause the heap to become fragmented.
  • Collection of the humongous objects is not as optimized by the G1 as with regular objects. It was especially troublesome with early Java 8 releases – until Java 1.8u40 the reclamation of humongous regions was only done during full GC events. More recent releases of the Hotspot JVM free the humongous regions at the end of the marking cycle during the cleanup phase, so the impact of the issue has been reduced significantly for newer JVMs.

To check whether or not your application is allocating objects in humongous regions, the first step would be to turn on GC logs similar to the following:

java -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintReferenceGC -XX:+UseG1GC -XX:+PrintAdaptiveSizePolicy -Xmx128m MyClass

Now, when you check the logs and discover sections like these:

 0.106: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: occupancy higher than threshold, occupancy: 60817408 bytes, allocation request: 1048592 bytes, threshold: 60397965 bytes (45.00 %), source: concurrent humongous allocation]
 0.106: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
 0.106: [G1Ergonomics (Concurrent Cycles) initiate concurrent cycle, reason: concurrent cycle initiation requested]
 0.106: [GC pause (G1 Humongous Allocation) (young) (initial-mark) 0.106: [G1Ergonomics (CSet Construction) start choosing CSet, _pending_cards: 0, predicted base time: 10.00 ms, remaining time: 190.00 ms, target pause time: 200.00 ms]

you have evidence that the application is indeed allocating humongous objects. The evidence is visible in the cause for a GC pause being identified as G1 Humongous Allocation and in the “allocation request: 1048592 bytes” section, where we can see that the application is trying to allocate an object with the size of 1,048,592 bytes, which is 16 bytes larger than the 50% of the 2 MB size of the humongous region specified for the JVM.

The first solution for humongous allocation is to change the region size so that (most) of the allocations would not exceed the 50% limit triggering allocations in the humongous regions. The region size is calculated by the JVM during startup based on the size of the heap. You can override the size by specifying -XX:G1HeapRegionSize=XX in the startup script. The specified region size must be between 1 and 32 megabytes and has to be a power of two.

This solution can have side effects – increasing the region size reduces the number of regions available so you need to be careful and run extra set of tests to see whether or not you actually improved the throughput or latency of the application.

A more time-consuming but potentially better solution would be to understand whether or not the application can limit the size of allocations. The best tools for the job in this case are profilers. They can give you information about humongous objects by showing their allocation sources with stack traces.

Conclusion

With the enormous number of possible applications that one may run on the JVM, coupled with the hundreds of JVM configuration parameters that one may tweak for GC, there are astoundingly many ways in which the GC may impact your application’s performance.

Therefore, there is no real silver bullet approach to tuning the JVM to match the performance goals you have to fulfill. What we have tried to do here is walk you through some common (and not so common) examples to give you a general idea of how problems like these can be approached. Coupled with the tooling overview and with a solid understanding of how the GC works, you have all the chances of successfully tuning garbage collection to boost the performance of your application.