Benchmarks
Introduction
Benchmark results are always dependent on a very individual setup. Normally it is not useful to generalize such results for every use case, but it can give you a hint. However, if you're really in the need of maximum performance, you should probably create an own benchmark with your objects or even use a profiler to detect the real hot spots in your application.
XStream uses the Java Microbenchmark Harness (JMH) of the JDK Tools as benchmark framework starting with version 1.4.9. As result it contains a ZIP file (xstream-jmh-<version>-app.zip) as new artifact containing anything required to run the benchmarks. Unpack the file and use the scripts in the bin directory to execute the benchmarks. Use option -h to look at the options provided by JMH. You may exchange the libraries in the lib directory with other versions of XStream or the individual parsers or you may even add new JMH benchmarks to the default ones of XStream.
All benchmark values below measure the average throughput in nanosecond per operation. JMH provides additional measurement options, see online help. The maximum deviation for each benchmark is recorded in the reference files of the distributed ZIP file. The benchmark is executed on Linux 4.1.12 Gentoo 64-bit system with an Intel Core i7 CPU 920 of 2.67 GHz. Note again, that these values are no replacement for real profiler results and they may vary from run to run (see reference files) due to this machine's background processes on a single CPU. However, it can give you some idea of what you can expect using different parser technologies.
Parser Benchmark
The values represent the average throughput of 15 runs with a single thread. The benchmarks emphasis the parser efficiency for different structures.
| Parser | Text | Array | Nested |
|---|---|---|---|
| W3C DOM (Oracle JDK 1.8.0_66) | 10037380.795 | 54234293.351 | 1917332.056 |
| JDOM (1.1.3) | 6368317.636 | 7910979.223 | 3862796.027 |
| JDOM 2 (2.0.5) | 5767640.105 | 10570210.653 | 2980906.727 |
| DOM4J (1.6.1) | 7816280.084 | 92998322.952 | 2108075.646 |
| XOM (1.1) | 7950778.533 | 38704485.310 | 2471019.743 |
| StAX (BEA 1.2.0) | 3108517.699 | 1310406.961 | 669111.164 |
| StAX (Woodstox 3.2.7) | 1884858.525 | 1240767.393 | 650470.623 |
| StAX (Oracle JDK 1.8.0_66) | 7366387.272 | 1334398.501 | 688229.709 |
| XPP (Xpp3 min 1.1.4c) | 2109341.076 | 1309607.210 | 3301732.767 |
| XPP (kXML2 min 2.3.0) | 3391204.266 | 1514514.680 | 8105934.241 |
| Binary (XStream 1.4.9) | 1144243.750 | 1062031.901 | 496839.565 |
| Jettison (1.2) | 3002547.220 | 1159238.555 | 682182.733 |
- Text
- A single element with a text of 100.000 characters.
- Array
- A single element with 1.000 child elements.
- Nested
- Nested elements in 500 levels.
Converter Type Benchmark
The values represent the average throughput of 16 runs with four threads using the Xpp3 parser for a structure with 1.000 elements. The benchmarks demonstrate the different converter types that can be used for a standard Java class.
| Converter Type | Throughput |
|---|---|
| Custom | 11276718.384 |
| Java Bean | 28878706.293 |
| Reflection | 40085786.696 |
- Custom
- A converter especially written for the Java type to convert.
- Java Bean
- Usage of the generic JavaBeanConverter, since the Java type respects the Java Bean contract.
- Reflection
- Usage of the generic converter based on reflection.
String Converter Benchmark
The values represent the average throughput of 16 runs with four threads using the Xpp3 parser for a structure with 10.000 string elements of various sizes and duplicates. The benchmarks demonstrate different implementations and configurations of the StringConverter.
| StringConverter Strategy | Throughput |
|---|---|
| No Cache | 19626160.696 |
| Intern | 23060982.052 |
| ConcurrentMap (length limit) | 21796001.29 |
| ConcurrentMap (unlimited) | 21378299.003 |
| Sync'd WeakCache (length limit) | 21838410.801 |
| Sync'd WeakCache (unlimited) | 22011251.691 |
- No Cache
- An implementation that does not cache deserialized String values with the consequence that repeated values will always allocate separate memory.
- Intern
- An implementation that uses String.intern() to cache the individual values. The memory pool used for the values is dependent on the JDK version. Up to Java 7 this was the permanent generation space i.e. the memory has to be shared with all loaded classes. It is up to the garbage collection when these string values are freed again.
- ConcurrentMap (length limit)
- An implementation that uses a ConcurrentHashMap as cache for strings of limited length (38 characters). The lifetime of the cache is equivalent with the lifetime of the XStream instance.
- ConcurrentMap (unlimited)
- An implementation that uses a ConcurrentHashMap as cache for all strings. The lifetime of the cache is equivalent with the lifetime of the XStream instance.
- Sync'd WeakCache (length limit)
- An implementation that uses a WeakCache for strings of limited length (38 characters). This cache uses weak references for its keys and values. An entry is therefore only kept as long as the deserialized object structure is referencing it. This is XStream's default strategy.
- Syn'd WeakCache (unlimited)
- An implementation that uses a WeakCache for all strings. This cache uses weak references for its keys and values. An entry is therefore ony kept as long as the deserialized object structure is referencing it.
Name Coder Benchmark
The values represent the average throughput of 25 runs with four threads using the Xpp3 parser for a structure with 250 nested elements using names invalid for XML elements. The benchmarks demonstrate different implementation strategies for a NameCoder to create valid tag names in XML.
| NameCoder Implementation | Throughput |
|---|---|
| No Coding | 7299475.975 |
| Dollar Coding | 7767196.902 |
| Escaped Underscore Coding | 9894393.289 |
| Cached Escaped Underscore Coding | 7867671.962 |
| Xml Friendly Coding | 8635501.208 |
- No Coding
- An implementation that does not encode the names of XML elements. It relies on the fact that the object graph does not contain elements with invalid XML names, because the name of the class types and members are either conforming or have been aliased.
- Dollar Coding
- An implementation that uses String.replace to replace any dollar sign with '·' (middle dot), a valid character normally not used for Java identifiers. In typical Java code there are no other invalid characters used for Java identifiers, however, Java allows identifiers to contain a wide range of UTF-8 characters and the JVM has even less restrictions.
- Escaped Underscore Coding
- An implementation that uses a StringBuilder to create the XML name by replacing any dollar sign with '_-' and escapes every plain underscore with two ones. The implementation will therefore only use characters in the standard ASCII range. It is not possible to use a simple minus sign as replacement because it is not a valid first character for XML names. However, the comments about Java identifiers and JVM identifiers from the Dollar Coding still apply.
- Cached Escaped Underscore Coding
- An implementation that implements a cache for the NameCoder that escapes the underscores.
- Xml Friendly Coding
- The default implementation of XStream using a StringBuilder and a cache, encoding any character that is invalid for XML names. It implements also the underscore escaping for compatibility reasons with XML created by earlier versions of XStream.