Running JVM-based Apps

GraalVM gives you an enhanced performance for JVM-based applications written in languages such as Java, Scala, Groovy, Clojure or Kotlin. It uses a dynamic compiler that dramatically improves the efficiency and the speed of applications through unique approaches to code analysis and optimization.

Interoperability

In addition to running JVM-based languages, you can also call any other language implemented with GraalVM Language Implementation Framework directly from Java. See the Polyglot Reference and the Embedding documentation for more information about interoperability with other programming languages.

Compiler Performance

The GraalVM compiler achieves excellent performance for modern workloads such as Scala or usage of the Java Streams API. For examples, see the Java performance examples.

Compiler Configuration on JVM

The options for configuring the GraalVM compiler on the JVM are in 3 categories.

General options

These are general options for setting/getting configuration details.

  • -XX:+UseJVMCICompiler: This disables use of the GraalVM compiler as the top tier JIT. This is useful when wanting to compare performance of the GraalVM compiler against the native JIT compilers.
  • -Dgraal.CompilerConfiguration=<name>: Selects the GraalVM compiler configuration to use. If omitted, the compiler configuration with the highest auto-selection priority is used. To see the set of available configurations, supply the value help to this option.

    The current configurations and their semantics are:

    • enterprise: To produce highly optimized code with a possible trade-off to compilation time. This value is only available in GraalVM EE.
    • community: To produce reasonably optimized code with a faster compilation time.
    • economy: To compile as fast as possible with less optimal throughput of the generated code.
  • -Dgraal.ShowConfiguration=none: Prints information about the GraalVM compiler configuration selected. This option only produces output when the compiler is initialized. By default, the GraalVM compiler is initialized on the first top-tier compilation. For this reason, the way to use this option is as follows: java -XX:+EagerJVMCI -Dgraal.ShowConfiguration=info -version.

    The accepted values for this option are:

    • none: To show no information.
    • info: To print one line of output showing the name of the compiler configuration in use and the location it is loaded from.
    • verbose: To print detailed compiler configuration information.
  • -Dgraal.MitigateSpeculativeExecutionAttacks=None: Selects a strategy to mitigate speculative execution attacks (e.g., SPECTRE).

    Accepted values are:

    • None: No mitigations are used in JIT compiled code.
    • AllTargets: All branches are protected against speculative attacks. This has a large performance impact.
    • GuardTargets: Only branches that preserve Java memory safety are protected. This has reduced performance impact.
    • NonDeoptGuardTargets: Same as GuardTargets except that branches which deoptimize are not protected since they can not be executed repeatedly.
  • --engine.Mode=default: Configures the execution mode of the engine. The execution mode automatically tunes the polyglot engine towards latency or throughput.
    • throughput: To collect the maximum amount of profiling information and compile using the maximum number of optimizations. This mode results in slower application startup but better throughput. This mode uses the compiler configuration community or enterprise if not specified otherwise.
    • default: To use a balanced engine configuration. This mode uses the compiler configuration community or enterprise if not specified otherwise.
    • latency: To collect only minimal profiling information and compile as fast as possible with less optimal generated code. This mode results in faster application startup but less optimal throughput. This mode uses the compiler configuration economy if not specified otherwise.
  • -XX:-UseJVMCINative Library: Disables use of libgraal. Instead, the GraalVM compiler is executed as classes loaded into the HotSpot heap. This will delay the time to reach peak performance as the compiler itself needs to be compiled before it produces code quickly. This mode allows you to debug the GraalVM compiler with a Java debugger.

Performance tuning options

  • -Dgraal.UsePriorityInlining=true: This can be used to disable use of the advanced inlining algorithm that favors throughput over compilation speed. This option is only available in GraalVM EE.
  • -Dgraal.Vectorization=true: This can be used to disable the auto vectorization optimization. This option is only available in GraalVM EE.
  • -Dgraal.TraceInlining=false: Enables tracing of inlining decisions. This can be used for advanced tuning where it may be possible to change the source code of the program. The output format is shown below:

    compilation of 'Signature of the compilation root method':
    at 'Sig of the root method' ['Bytecode index']: <'Phase'> 'Child method signature': 'Decision made about this callsite'
      at 'Signature of the child method' ['Bytecode index']:
         |--<'Phase 1'> 'Grandchild method signature': 'First decision made about this callsite'
         \--<'Phase 2'> 'Grandchild method signature': 'Second decision made about this callsite'
      at 'Signature of the child method' ['Bytecode index']: <'Phase'> 'Another grandchild method signature': 'The only decision made about this callsite.'
    

    For example:

    compilation of java.lang.Character.toUpperCaseEx(int):
    at java.lang.Character.toUpperCaseEx(Character.java:7138) [bci: 22]:
       ├──<GraphBuilderPhase> java.lang.CharacterData.of(int): no, bytecode parser did not replace invoke
       └──<PriorityInliningPhase> java.lang.CharacterData.of(int): yes, worth inlining according to the cost-benefit analysis.
    at java.lang.Character.toUpperCaseEx(Character.java:7138) [bci: 26]:
       ├──<GraphBuilderPhase> java.lang.CharacterDataLatin1.toUpperCaseEx(int): no, bytecode parser did not replace invoke
       └──<PriorityInliningPhase> java.lang.CharacterDataLatin1.toUpperCaseEx(int): yes, worth inlining according to the cost-benefit analysis.
      at java.lang.CharacterDataLatin1.toUpperCaseEx(CharacterDataLatin1.java:223) [bci: 4]:
         ├──<GraphBuilderPhase> java.lang.CharacterDataLatin1.getProperties(int): no, bytecode parser did not replace invoke
         └──<PriorityInliningPhase> java.lang.CharacterDataLatin1.getProperties(int): yes, worth inlining according to the cost-benefit analysis.
    

Diagnostic options

  • -Dgraal.CompilationFailureAction=Silent: Specifies the action to take when compilation fails by throwing an exception.

    The accepted values are:

    • Silent: Print nothing to the console.
    • Print: Print a stack trace to the console.
    • Diagnose: Retry the compilation with extra diagnostics enabled. On VM exit, the collected diagnostics are saved to a zip file that can be submitted along with a bug report. A message is printed to the console describing where the diagnostics file is saved:
      Graal diagnostic output saved in /Users/graal/graal_dumps/1549459528316/graal_diagnostics_22774.zip
      
    • ExitVM: Same as Diagnose except that the VM process exits after retrying.

    For all values except for ExitVM, the VM continues executing.

  • -Dgraal.CompilationBailoutAsFailure=false: The compiler may not complete compilation of a method due to some property or code shape in the method (e.g. exotic uses of the jsr and ret bytecodes). In this case the compilation bails out. If you want to be informed of such bailouts, this option makes GraalVM treat bailouts as failures and thus be subject to the action specified by the -Dgraal.CompilationFailureAction option.
  • -Dgraal.PrintCompilation=false: Prints an informational line to the console for each completed compilation. For example:
    HotSpotCompilation-11  Ljava/lang/Object;                            wait          ()V       |  591ms    12B    92B  4371kB
    HotSpotCompilation-175 Ljava/lang/String;                            lastIndexOf   (II)I     |  590ms   126B   309B  4076kB
    HotSpotCompilation-184 Ljava/util/concurrent/ConcurrentHashMap;      setTabAt      ([Ljava/util/concurrent/ConcurrentHashMap$Node;ILjava/util/concurrent/ConcurrentHashMap$Node;)V  |  591ms    38B    67B  3411kB
    HotSpotCompilation-136 Lsun/nio/cs/UTF_8$Encoder;                    encode        ([CII[B)I |  591ms   740B   418B  4921
    

Setting Compiler Options with Language Launchers

The GraalVM compiler properties above are usable with some other GraalVM launchers such as node, js and lli. The prefix for specifying the properties is slightly different. For example:

$ java -XX:+EagerJVMCI -Dgraal.ShowConfiguration=info -version

Becomes:

$ js --jvm --vm.Dgraal.ShowConfiguration=info -version

Note the -D prefix is replaced by --vm.D.

JVM Operations Manual

Difference between running the GraalVM compiler in a Native Image vs on the JVM

When running the GraalVM compiler on the JVM, it goes through the same warmup phase that the rest of Java application does. That is, it is first interpreted before its hot methods are compiled. This can translate into slightly longer times until the application reaches peak performance when compared to the native compilers in the JVM such as C1 and C2.

To address the issue of taking longer to reach to peak performance, both GraalVM Enterprise and Community images now include libgraal – a shared library, produced using Native Image framework to ahead-of-time compile the compiler itself. That means the GraalVM compiler can now be deployed as a native shared library. In this mode, the compiler uses memory separate from the HotSpot heap and it runs compiled from the start. That is, it has execution properties similar to other native HotSpot compilers such as C1 and C2. Currently, this is the default mode of operation. It can be disabled with -XX:-UseJVMCINativeLibrary.

Measuring Performance

The first thing to be sure of when measuring the performance is to ensure the JVM is using the GraalVM compiler. In the GraalVM binary, the JVM is configured to use the GraalVM compiler as the top tier compiler by default. You can confirm this by adding -Dgraal.ShowConfiguration=info to the command line. It will produce a line of output similar to the one below when the compiler is initialized:

Using Graal compiler configuration 'community' provided by org.graalvm.compiler.hotspot.CommunityCompilerConfigurationFactory loaded from jar:file:/Users/dsimon/graal/graal/compiler/mxbuild/dists/graal.jar!/org/graalvm/compiler/hotspot/CommunityCompilerConfigurationFactory.class

Note that the GraalVM compiler is only initialized on the first top tier JIT compilation request so if your application is short lived, you may not see this output.

Optimizing JVM-based applications is a science in itself. The compilation may not even be a factor in the case of poor performance as the problem may lie in any other part of the VM (I/O, garbage collection, threading etc) or in poorly written application or 3rd party library code. For this reason, it’s worth using profilers such as Java Mission Control to diagnose application behavior.

You can also compare performance against the native top tier compiler in the JVM by adding -XX:-UseJVMCICompiler to the command line.

If you observe a significant performance regression when using the GraalVM compiler, please open an issue on GitHub. Attaching a Java Flight Recorder log and instructions to reproduce the issue makes investigation easier and thus chances of a fix higher. Even better is if you can submit a JMH benchmark that represents the hottest parts of your application (as identified by a profiler). This allows us to very quickly pinpoint missing optimization opportunities or to offer suggestions on how to restructure the code to avoid or reduce performance bottlenecks.

Troubleshooting the GraalVM Compiler

Like all software, the GraalVM compiler is not guaranteed to be bug free so it is useful to know how to diagnose and submit useful bug reports if you encounter issues.

If you spot a security vulnerability, please do not report it via GitHub Issues or the public mailing lists, but via the process outlined at Reporting Vulnerabilities guide.

Compilation Exceptions

One advantage of the compiler being written in Java is that runtime exceptions during compilation are not fatal VM errors. Instead, each compilation has an exception handler that takes an action based on the graal.CompilationFailureAction property.

The default value is Silent. Specifying Diagnose causes failing compilations to be retried with extra diagnostics enabled. Just before the VM exits, all diagnostic output captured during retried compilations is written to a .zip file and its location is printed on the console:

Graal diagnostic output saved in /Users/demo/graal-dumps/1499768882600/graal_diagnostics_64565.zip

You can then attach the .zip file to an issue on GitHub.

Apart from Silent and Diagnose the following values for graal.CompilationFailureAction are supported:

  • Print: Prints a message and stack trace to the console but does not do the re-compilation.
  • ExitVM: Same as Diagnose but the VM process exits after the re-compilation.

Code Generation Errors

The other type of error compilers can have is producing incorrect machine code. This error can cause a VM crash, which should produce a file that starts with hs_err_pid in the current working directory of the VM process. In most cases, there is a section in the file that shows the stack at the time of the crash, including the type of code for each frame in the stack, as in the following example:

Stack: [0x00007000020b1000,0x00007000021b1000],  sp=0x00007000021af7a0,  free space=1017k
Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native code)
J 761 JVMCI org.graalvm.compiler.core.gen.NodeLIRBuilder.matchComplexExpressions(Ljava/util/List;)V (299 bytes) @ 0x0000000108a2fc01 [0x0000000108a2fac0+0x141] (null)
j  org.graalvm.compiler.core.gen.NodeLIRBuilder.doBlock(Lorg/graalvm/compiler/nodes/cfg/Block;Lorg/graalvm/compiler/nodes/StructuredGraph;Lorg/graalvm/compiler/core/common/cfg/BlockMap;)V+211
j  org.graalvm.compiler.core.LIRGenerationPhase.emitBlock(Lorg/graalvm/compiler/nodes/spi/NodeLIRBuilderTool;Lorg/graalvm/compiler/lir/gen/LIRGenerationResult;Lorg/graalvm/compiler/nodes/cfg/Block;Lorg/graalvm/compiler/nodes/StructuredGraph;Lorg/graalvm/compiler/core/common/cfg/BlockMap;)V+65

This example shows that the top frame was compiled (J) by the JVMCI compiler, which is the GraalVM compiler. The crash occurred at offset 0x141 in the machine code produced for:

org.graalvm.compiler.core.gen.NodeLIRBuilder.matchComplexExpressions(Ljava/util/List;)V

The next two frames in the stack were executing in the interpreter (j). The location of the crash is also often indicated near the top of the file with something like this:

# Problematic frame:
# J 761 JVMCI org.graalvm.compiler.core.gen.NodeLIRBuilder.matchComplexExpressions(Ljava/util/List;)V (299 bytes) @ 0x0000000108a2fc01 [0x0000000108a2fac0+0x141] (null)

In this example, there is likely an error in the code produced by the GraalVM compiler for NodeLIRBuilder.matchComplexExpressions.

When filing an issue on GitHub for such a crash, you should first attempt to reproduce the crash with extra diagnostics enabled for the compilation of the problematic method. In this example, you would add the following to your command line:

-Dgraal.MethodFilter=NodeLIRBuilder.matchComplexExpressions, -Dgraal.Dump=:2

These options are described in more detail here. In brief, these options tell the compiler to capture snapshots of the compiler state at verbosity level 2 while compiling any method named matchComplexExpressions in a class with a simple name of NodeLIRBuilder. The complete format of the MethodFilter option is described in the output of java -XX:+JVMCIPrintProperties.

Quite often, the crash location does not exist directly in the problematic method mentioned in the crash log but comes from an inlined method.

In such a case, simply filtering for the problematic method might not capture an erroneous compilation causing a crash.

To improve the likelihood of capturing an erroneous compilation, you need to broaden the MethodFilter value. To guide this, add -Dgraal.PrintCompilation=true when trying to reproduce the crash so you can see what was compiled just before the crash.

The following shows sample output from the console:

HotSpotCompilation-1218        Lorg/graalvm/compiler/core/amd64/AMD64NodeLIRBuilder;                  peephole                                      (Lorg/graalvm/compiler/nodes/ValueNode;)Z           |   87ms   428B   447B  1834kB
HotSpotCompilation-1212        Lorg/graalvm/compiler/lir/LIRInstructionClass;                         forEachState                                  (Lorg/graalvm/compiler/lir/LIRInstruction;Lorg/graalvm/compiler/lir/InstructionValueProcedure;)V  |  359ms    92B   309B  6609kB
HotSpotCompilation-1221        Lorg/graalvm/compiler/hotspot/amd64/AMD64HotSpotLIRGenerator;          getResult                                     ()Lorg/graalvm/compiler/hotspot/HotSpotLIRGenerationResult;  |   54ms    18B   142B  1025kB
#
# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0x000000010a6cafb1, pid=89745, tid=0x0000000000004b03
#
# JRE version: OpenJDK Runtime Environment (8.0_121-b13) (build 1.8.0_121-graalvm-olabs-b13)
# Java VM: OpenJDK 64-Bit GraalVM (25.71-b01-internal-jvmci-0.30 mixed mode bsd-amd64 compressed oops)
# Problematic frame:
# J 1221 JVMCI org.graalvm.compiler.hotspot.amd64.AMD64HotSpotLIRGenerator.getResult()Lorg/graalvm/compiler/hotspot/HotSpotLIRGenerationResult; (18 bytes) @ 0x000000010a6cafb1 [0x000000010a6caf60+0x51] (null)
#
# Failed to write core dump. Core dumps have been disabled. To enable core dumping, try "ulimit -c unlimited" before starting Java again

Here we see that the crash happened in a different method than the first crash. As such, we expand the filter argument to be -Dgraal.MethodFilter= NodeLIRBuilder.matchComplexExpressions,AMD64HotSpotLIRGenerator.getResult and run again.

When the VM crashes in this way, it does not execute the shutdown code that archives the GraalVM compiler diagnostic output or delete the directory it was written to. This must be done manually after the crash.

By default, the directory is $PWD/graal-dumps/<timestamp>; for example, ./graal-dumps/1499938817387. However, you can set the directory with -Dgraal.DumpPath=<path>.

A message, such as the following, is printed to the console when this directory is first used by the compiler:

Dumping debug output in /Users/demo/graal-dumps/1499768882600

This directory should contain content related to the crashing method, such as:

$ ls -l /Users/demo/graal-dumps/1499768882600
-rw-r--r--  1 demo  staff    144384 Jul 13 11:46 HotSpotCompilation-1162[AMD64HotSpotLIRGenerator.getResult()].bgv
-rw-r--r--  1 demo  staff     96925 Jul 13 11:46 HotSpotCompilation-1162[AMD64HotSpotLIRGenerator.getResult()].cfg
-rw-r--r--  1 demo  staff  12600725 Jul 13 11:46 HotSpotCompilation-791[NodeLIRBuilder.matchComplexExpressions(List)].bgv
-rw-r--r--  1 demo  staff   1727409 Jul 13 11:46 HotSpotCompilation-791[NodeLIRBuilder.matchComplexExpressions(List)].cfg

You should attach a zip of this directory to an issue on GitHub.