Skip to content

Java Agent for Memory Measurements

License

Notifications You must be signed in to change notification settings

clojure-goes-fast/jamm

 
 

Repository files navigation

Overview

Jamm provides MemoryMeter, a Java agent for all Java versions to measure actual object memory use including JVM overhead.

Use

To use MemoryMeter, start the JVM with -javaagent:<path to>/jamm.jar.

You can then use MemoryMeter in your code like this:

MemoryMeter meter = MemoryMeter.builder().build();
meter.measure(object);
meter.measureDeep(object);

See MemoryMeter.Builder for more options.

If you would like to use MemoryMeter in a web application, make sure that you do NOT put this jar in WEB-INF/lib, as that may cause problems since your code is accessing a MemoryMeter from a different class loader than the one loaded by the -javaagent and won't see it as initialized.

If you want MemoryMeter not to measure some specific classes, you can mark the classes (or interfaces) using the @Unmetered annotation.

It is good to reuse existing MemoryMeter instances. Creating new MemoryMeter instances can and will cause significant performance penalties and also unnecessary side effects, since all guess-modes rely on java.lang.ClassValue.

Dependency coordinates for the latest version of Jamm

Gradle:

dependencies {
    implementation("com.github.jbellis:jamm:0.4.0")
}

Maven:

<dependency>
    <groupId>com.github.jbellis</groupId>
    <artifactId>jamm</artifactId>
    <version>0.4.0</version>
</dependency>

Building

Build + test:

./gradlew jar test

To run the microbenchmark:

./gradlew microbench
build/microbench <JMH options...>

Public to local Maven repo (in case you need that)

./gradlew publishToMavenLocal

IDEs

You can just open the project in IntelliJ. If you've opened the Maven-ish branch(es) in IntelliJ before, you'll have to remove the .idea folder.

Benchmarks

See Microbench.java

To run the microbenchmarks, execute ./gradlew microbench and run build/microbench from the command line.

Some microbenchmark results (JVM options: -Xms16g -Xmx16g -XX:+UseG1GC -XX:+AlwaysPreTouch):

Java 11.0.8

Benchmark                                 (guess)  (nested)  (refs)  Mode  Cnt   Score   Error  Units
Microbench.arrayByteArray             ALWAYS_SPEC       100       4  avgt    3   0.012 ± 0.019  us/op
Microbench.arrayByteArray           ALWAYS_UNSAFE       100       4  avgt    3   0.012 ± 0.019  us/op
Microbench.arrayByteArray  ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.012 ± 0.017  us/op
Microbench.cls1                       ALWAYS_SPEC       100       4  avgt    3   0.026 ± 0.004  us/op
Microbench.cls1                     ALWAYS_UNSAFE       100       4  avgt    3   0.026 ± 0.005  us/op
Microbench.cls1            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.026 ± 0.004  us/op
Microbench.cls2                       ALWAYS_SPEC       100       4  avgt    3   0.350 ± 0.155  us/op
Microbench.cls2                     ALWAYS_UNSAFE       100       4  avgt    3   0.387 ± 0.030  us/op
Microbench.cls2            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.355 ± 0.040  us/op
Microbench.cls3                       ALWAYS_SPEC       100       4  avgt    3   0.761 ± 0.087  us/op
Microbench.cls3                     ALWAYS_UNSAFE       100       4  avgt    3   0.733 ± 0.130  us/op
Microbench.cls3            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.738 ± 0.099  us/op
Microbench.deeplyNested               ALWAYS_SPEC       100       4  avgt    3  49.825 ± 1.482  us/op
Microbench.deeplyNested             ALWAYS_UNSAFE       100       4  avgt    3  46.260 ± 4.466  us/op
Microbench.deeplyNested    ALWAYS_INSTRUMENTATION       100       4  avgt    3  52.301 ± 8.285  us/op
Microbench.justByteArray              ALWAYS_SPEC       100       4  avgt    3   0.014 ± 0.005  us/op
Microbench.justByteArray            ALWAYS_UNSAFE       100       4  avgt    3   0.014 ± 0.004  us/op
Microbench.justByteArray   ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.014 ± 0.006  us/op
Microbench.justByteBuffer             ALWAYS_SPEC       100       4  avgt    3   0.056 ± 0.011  us/op
Microbench.justByteBuffer           ALWAYS_UNSAFE       100       4  avgt    3   0.058 ± 0.018  us/op
Microbench.justByteBuffer  ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.058 ± 0.029  us/op
Microbench.justString                 ALWAYS_SPEC       100       4  avgt    3   0.059 ± 0.061  us/op
Microbench.justString               ALWAYS_UNSAFE       100       4  avgt    3   0.057 ± 0.019  us/op
Microbench.justString      ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.056 ± 0.001  us/op

Java 8u265

Benchmark                                 (guess)  (nested)  (refs)  Mode  Cnt   Score    Error  Units
Microbench.arrayByteArray             ALWAYS_SPEC       100       4  avgt    3   0.011 ±  0.007  us/op
Microbench.arrayByteArray           ALWAYS_UNSAFE       100       4  avgt    3   0.011 ±  0.006  us/op
Microbench.arrayByteArray  ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.011 ±  0.007  us/op
Microbench.cls1                       ALWAYS_SPEC       100       4  avgt    3   0.023 ±  0.009  us/op
Microbench.cls1                     ALWAYS_UNSAFE       100       4  avgt    3   0.025 ±  0.039  us/op
Microbench.cls1            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.024 ±  0.012  us/op
Microbench.cls2                       ALWAYS_SPEC       100       4  avgt    3   0.357 ±  0.034  us/op
Microbench.cls2                     ALWAYS_UNSAFE       100       4  avgt    3   0.365 ±  0.008  us/op
Microbench.cls2            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.366 ±  0.100  us/op
Microbench.cls3                       ALWAYS_SPEC       100       4  avgt    3   0.773 ±  0.178  us/op
Microbench.cls3                     ALWAYS_UNSAFE       100       4  avgt    3   0.754 ±  0.328  us/op
Microbench.cls3            ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.755 ±  0.396  us/op
Microbench.deeplyNested               ALWAYS_SPEC       100       4  avgt    3  46.889 ± 18.817  us/op
Microbench.deeplyNested             ALWAYS_UNSAFE       100       4  avgt    3  45.166 ±  9.059  us/op
Microbench.deeplyNested    ALWAYS_INSTRUMENTATION       100       4  avgt    3  48.041 ± 14.777  us/op
Microbench.justByteArray              ALWAYS_SPEC       100       4  avgt    3   0.011 ±  0.008  us/op
Microbench.justByteArray            ALWAYS_UNSAFE       100       4  avgt    3   0.011 ±  0.009  us/op
Microbench.justByteArray   ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.011 ±  0.007  us/op
Microbench.justByteBuffer             ALWAYS_SPEC       100       4  avgt    3   0.053 ±  0.013  us/op
Microbench.justByteBuffer           ALWAYS_UNSAFE       100       4  avgt    3   0.053 ±  0.011  us/op
Microbench.justByteBuffer  ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.053 ±  0.014  us/op
Microbench.justString                 ALWAYS_SPEC       100       4  avgt    3   0.055 ±  0.010  us/op
Microbench.justString               ALWAYS_UNSAFE       100       4  avgt    3   0.054 ±  0.007  us/op
Microbench.justString      ALWAYS_INSTRUMENTATION       100       4  avgt    3   0.054 ±  0.019  us/op

Previous (0.3.x) versions of jamm (w/ 8u265)

Mentioned for comparison reasons! Performance numbers for the recent version above.

Benchmark                        (guess)  (nested)  (refs)  Mode  Cnt    Score    Error  Units
Microbench.cls1              ALWAYS_SPEC       100       4  avgt    3    0.243 ±  0.170  us/op
Microbench.cls1            ALWAYS_UNSAFE       100       4  avgt    3    0.304 ±  0.078  us/op
Microbench.cls1                    NEVER       100       4  avgt    3    0.210 ±  0.126  us/op
Microbench.cls2              ALWAYS_SPEC       100       4  avgt    3    2.924 ±  0.272  us/op
Microbench.cls2            ALWAYS_UNSAFE       100       4  avgt    3    3.086 ±  0.419  us/op
Microbench.cls2                    NEVER       100       4  avgt    3    2.624 ±  0.131  us/op
Microbench.cls3              ALWAYS_SPEC       100       4  avgt    3    5.919 ±  1.433  us/op
Microbench.cls3            ALWAYS_UNSAFE       100       4  avgt    3    6.031 ±  1.632  us/op
Microbench.cls3                    NEVER       100       4  avgt    3    5.329 ±  0.513  us/op
Microbench.deeplyNested      ALWAYS_SPEC       100       4  avgt    3  500.569 ± 27.362  us/op
Microbench.deeplyNested    ALWAYS_UNSAFE       100       4  avgt    3  495.529 ± 66.984  us/op
Microbench.deeplyNested            NEVER       100       4  avgt    3  388.130 ± 83.949  us/op
Microbench.justByteArray     ALWAYS_SPEC       100       4  avgt    3    0.012 ±  0.011  us/op
Microbench.justByteArray   ALWAYS_UNSAFE       100       4  avgt    3    0.013 ±  0.004  us/op
Microbench.justByteArray           NEVER       100       4  avgt    3    0.055 ±  0.010  us/op
Microbench.justByteBuffer    ALWAYS_SPEC       100       4  avgt    3    0.942 ±  0.230  us/op
Microbench.justByteBuffer  ALWAYS_UNSAFE       100       4  avgt    3    0.935 ±  0.154  us/op
Microbench.justByteBuffer          NEVER       100       4  avgt    3    0.689 ±  0.247  us/op
Microbench.justString        ALWAYS_SPEC       100       4  avgt    3    0.616 ±  0.165  us/op
Microbench.justString      ALWAYS_UNSAFE       100       4  avgt    3    0.684 ±  0.636  us/op
Microbench.justString              NEVER       100       4  avgt    3    0.574 ±  0.098  us/op

The fine print

MemoryMeter can use java.lang.instrument.Instrumentation.getObjectSize(), which only claims to provide "approximate" results, but in practice seems to work as expected.

Two more implementations are available as well. One is using sun.misc.Unsafe w/ reflection and one just uses reflection. Both are inaccurate and rely on guesses. It is strongly recommended to not use "unsafe" or "specification" and jamm will emit a warning on stderr:

  • The "Unsafe" implementation performs arithmetics on the "cookies" returned by Unsafe.objectFieldOffset(), althought the Javadoc says: Do not expect to perform any sort of arithmetic on this offset; it is just a cookie which is passed to the unsafe heap memory accessors. The implementation does not always consider Java object layouts in under all circumstances for all JVMs.
  • The "specification" is just a best-effort guess-timate that can easily produce results that are wrong w/ newer JVMs and GCs and/or newer Java features.

0.4.0 notes

  • The method MemoryMeter.countChildren() has been removed, because there is a) no corresponding implementation from JDK-8249196 and b) there was no test coverage at all.
  • The construction of the MemoryMeter instance has been refactored to a proper builder pattern to allow distinct classes for each implementation: Instrumentation, Unsafe, Specification).
  • A bunch of internal optimizations went into the existing guess-modes and performance for these modes has also been significantly improved (about 10x, see microbenchmark results above).
  • Support for @Unmetered for fields has been removed for the future implementation of java.lang.Runtime.deepSizeOf() with Java 16.

About

Java Agent for Memory Measurements

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Java 90.6%
  • Kotlin 9.4%