Skip to content

Latest commit

 

History

History
304 lines (285 loc) · 27 KB

IoSerialiser.md

File metadata and controls

304 lines (285 loc) · 27 KB
Raw

#Yet - another - Serialiser(YaS)

# or : why we are not reusing... and opted to write yet another custom data serialiser

Data serialisation is a basic core functionality when information has to be transmitted, stored and later retrieved by (often quite) different subsystems. With a multitude of different serialiser libraries, a non-negligible subset of these claim to be the fastest, most efficient, easiest-to-use or < add your favourite superlative here >.

While this may be certainly true for the libraries' original design use-case, this often breaks down for other applications that are often quite diverse and may focus on different aspects depending on the application. Hence, a fair comparison of their performance is usually rather complex and highly non-trivial because the underlying assumptions of 'what counts as important' being quite different between specific domains, boundary constraints, application goals and resulting (de-)serialisation strategies. Rather than claiming any superlative, or needlessly bashing solutions that are well suited for their design use-cases, we wanted to document the considerations, constraints and application goals of our specific use-case and that guided our multi-protocol serialiser developments. This also in the hope that it might find interest, perhaps adoption, inspires new ideas, or any other form of improvements. Thus, if you find something missing, unclear, or things that could be improved, please feel encouraged to post a PR.

DISCLAIMER: This specific implementation while not necessarily a direct one-to-one source-code copy is at least conceptually based upon a combination of other open-sourced implementations, long-term experience with internal-proprietary wire-formats, and new serialiser design ideas expressed in the references below which were adopted, adapted and optimised for our specific use-case.

Our Use-Case

We use this and Chart-Fx in order to aid the development of functional microservices that monitor and control a large variety of device- and beam-based parameters that are necessary for the operation of our FAIR particle accelerators. These microservices cover in particular those that require the aggregation of measurement data from different sub-systems, or that require domain-specific logic or real-time signal-processing algorithms that cannot be efficiently implemented in any other single device or sub-system.

Quick Overview

This serialiser implementation defines three levels of interface abstractions:

  • IoBuffer which defines the low-level byte-array format of how data primitives (ie. boolean, byte, ...,float, double), String, and their array counter-part (ie. boolean[], byte[], ...,float[]', 'double[], String[]) are stored. There are two default implementations:
    • ByteBuffer which basically wraps around and extends java.nio.ByteBuffer to also support String and primitive arrays, and
    • FastByteBuffer which is the recommended (~ 25% faster) reimplementation using direct byte-array and cached field accesses.
  • IoSerialiser which defines the compound wire-format for more complex objects (e.g. List<T>, Map<K,V>, multi-dimensional arrays etc), including field headers, and annotations. There are three default implementations: (N.B. IoSerialiser allows further extensions to any other structurally similar protocol. A robust implementation of IoSerialiser::checkHeaderInfo() is critical in order to distinguish new protocols from existing ones.)
    • BinarySerialiser which is the primary binary-based transport protocol used by this library,
    • CmwLightSerialiser which is the backward compatible re-implementation of an existing proprietary protocol internally used in our facility, and
    • JsonSerialiser which implements the JSON protocol commonly used in RESTful HTTP-based services.
  • IoClassSerialiser which deals with the automatic mapping and (de-)serialisation between the class field structure and specific wire-format. This class defines default strategies for generic and nested classes and can be further extended by custom serialiser prototypes for more complex classes, other custom nested protocols or interfaces using the FieldSerialiser interface.

A short working example of how these can be used is shown in IoClassSerialiserSimpleTest:

@Test
void simpleTest() {
    final IoBuffer byteBuffer = new FastByteBuffer(10_000); // alt: new ByteBuffer(10_000);
    final IoClassSerialiser ioClassSerialiser = new IoClassSerialiser(byteBuffer, BinarySerialiser.class);
    TestDataClass data = new TestDataClass(); // object to be serialised

    byteBuffer.reset();
    ioClassSerialiser.serialiseObject(data); // pojo -> serialised data
    // [..] stream/write serialised byteBuffer content [..]

    // [..] stream/read serialised byteBuffer content
    byteBuffer.flip(); // mark byte-buffer for reading
    TestDataClass received = ioClassSerialiser.deserialiseObject(TestDataClass.class);

    // check data equality, etc...
    assertEquals(data, received);
}

The specific wire-format that the IoClassSerialiser uses can be set either programmatically or dynamically (auto-detection based on serialised data content header) via:

    ioClassSerialiser.setMatchedIoSerialiser(BinarySerialiser.class);
    ioClassSerialiser.setMatchedIoSerialiser(CmwLightSerialiser.class);
    ioClassSerialiser.setMatchedIoSerialiser(JsonSerialiser.class);
    // to auto-detect the suitable serialiser based on serialised data header:
    ioClassSerialiser.setAutoMatchSerialiser(true);

The extension for arbitrary custom classes or interfaces can be achieved through (here for the DoubleArrayList class) via:

    serialiser.addClassDefinition(new FieldSerialiser<>(
        (io, obj, field) -> field.getField().set(obj, DoubleArrayList.wrap(io.getDoubleArray())), // IoBuffer &rightarrow; class field reader function
        (io, obj, field) -> DoubleArrayList.wrap(io.getDoubleArray()), // return function - generates new object based on IoBuffer content
        (io, obj, field) -> { // class field &rightarrow; IoBuffer writer function
    final DoubleArrayList retVal = (DoubleArrayList) field.getField().get(obj);
    io.put(field, retVal.elements(), retVal.size());
        },
        DoubleArrayList.class));

The DataSetSerialiser serialiser implementation is a representative example and serialises the DataSet interface into an abstract implementation-independent wire-format using the FieldDataSetHelper function. This is also the most prominent common domain object definition that is used within our MVC-pattern driven microservice-, data-processing-, and UI-applications and one of the original primary motivations why we designed and built the IoClassSerialiser implementation.

Primary Serialiser Functionality Goals and Constraints

Some of the aspects that were incorporated into the design, loosely ordered according to their importance:

  1. performance: providing an optimised en-/decoding that minimises the effective total latency between the data object content being ready to be serialised and sent by the server until the object is received, fully de-serialised and ready for further processing on the client-side. N.B. some serialisers trade-off size for en-/decoding speed, which may be suitable for primarily network-io limited systems. Since io-bandwidth is not a primary concern for our local network, we chose a rather simple encoder with no explicit compression stage to save CPU clock cycles.
  2. facilitate multi-protocol implementations, protocol evolution and loose coupling between data object definitions on the server- and corresponding client-side, ie. services and clients may communicate with different protocol versions and need to agree only on a mandatory small sub-set of information they both need to share. N.B. most micro-services develop naturally and grow their functionality with time. This decoupling is necessary to provide a smooth and soft upgrade path for early adopters that require these new functionalities (ie. thus also being updated during regular facility operation), and clients that may require a controlled maintenance period, e.g. safety related systems, that need a formal qualification process prior to being deployed into regular operation with a modified data-model.
  3. same client- and server-side API, decoupling the serialisers' wire-formats (ie. different binary formats, JSON, XML, YML, ...) from the specific microservice APIs and low-level transport protocols that transmit the serialised data N.B. encapsulates domain-specific control as well as the generic microservice logic into reusable code blocks that can be re-implemented if deemed necessary, and that are decoupled from the specific required io-formats, which are usually either driven by technical necessity (e.g. device supporting only one data wire-format) and/or client-side preferences (e.g. web-based clients typically favouring RESTful JSON-based protocols while high-throughput clients with real-time requirements often favour more optimised binary data protocols over TCP/UDP-based sockets).
  4. derive schemas for generic data directly from C++ or Java class structures and basic types rather than a 3rd-party IDL definition (ie. using Pocos & Pojos as IDL)
    • aims at a high compatibility between C++, Java and other languages derived thereof and leverages existing experience of developers with those languages N.B. this improves the productivity of new/occasional/less-experienced users who need to be ony vaguely familiar with C++/Java and do not need to learn yet another new dedicated DSL. This also inverts the problem: rather than 'here are the data structures you allowed to use to be serialised' to 'what can be done to serialise the structures one already is using'.
    • enforces stronger type-safety N.B. some other serialisers encode only sub-sets of the possible data structures, and or reduce the specific type to encompassing super types. For example, integer-types such as byte, short, int all being mapped to long, or all floating-point-type numbers to double which due to the ambiguity causes unnecessary numerical decoding errors on the deserialisation side.
    • support for simple data primitives, nested class objects or common data container, such as Collection<T>, List<T>, Set<T>, ..., Map<K,V>, etc. N.B. We found, that due to the evolution of our microservices and data protocol definitions, we frequently had to remap and rewrite adapters between our internal map-based data-formats and class objects which proved to be a frequent and unnecessary source of coding errors.
    • efficient (first-class) support of large collections of numeric (floating-point) data N.B. many of the serialiser supporting binary wire-format seem to be optimised for simple data structure that are typically much smaller than 1k Bytes rather than large numeric arrays that were eiter slow to encode and/or required custom serialiser extensions.
    • usage of compile-time reflection ie. offline/one-time optimisation prior to running → run deterministic/optimally while online w/o relying on dynamic parsing optimisations N.B. this particularly simplifies the evolution, modification of data structures, and removes one of the common source of coding errors, since the synchronisation between class-structure, serialised-data-structure and formal-IDL-structure is omitted.
    • optional: support run-time reflection as a fall-back solution for new data/users that haven't used the compile-time reflection
    • optional support of UTF-8-based and fall-back to ISO8859-1-based String encoding if a faster or more efficient en-/decoding is needed.
  5. allow schema extensions through optional custom (de-)serialiser routines for known classes or interface that are either more optimised, or that implement a specific data-exchange format for a given generic class interface.
  6. self-documented data-structures with optional data field annotations to document and communicate the data-exchange-API-intend to the client
    • some examples: 'unit' and 'description' of specific data fields, read/write field access definitions, definition of field sub-sets, etc.
    • allows on-the-fly data structure (full schema) documentation for users based on the transmitted wire-format structure w/o the explicite need to have access to the exact service class domain object definition N.B. we keep the code public, this also facilitate automatic documentation updates whenever the code is being modified and opens the possibility of OpenAPI specification -style extensions common for RESTful service.
    • optional: full schema information is transmitted only for the first and (optionally) suppressed in subsequent transmissions for improved performance. N.B. trade-off between optimise latency/throughput in high-volume paths vs. feature-rich/documented data storage protocol for less critical low-volume 'get/set' operations.
  7. minimise code-base and code-bloat -- for two reasons:
    • smaller code usually leads to smaller compiled binary sizes that are more likely to fit into CPU cache, thus are less likely to be evicted on context changes, and result into overall faster code. N.B. while readability is an important issue, we found that certain needless use of 'interface + impl pattern' (ie. only one implementation for a given interface) are harder to read and harder to optimise for by the (JIT) compiler too. As an example, in-lining and keeping the code in one (albeit larger) source file proved to yield much faster results for the CmwLightSerialiser that is a reimplementation of an existing internally used wire-format.
    • maintenance: code should be able to be re-engineered or optimised within typically 2 weeks by one skilled developer. N.B. more code requires more time to read and to understand. While there are many skilled developer, having a simple code base also implies that the code can be more easily be modified, tested, fixed or maintained by any internally available developer. Also, we believe that this makes it possibly more likely to be adopted by external users that want to understand, upgrade, or bug-fix of 'what is under the hood' and is of specific interest to them. Having too many paradigms, patterns or library dependencies -- even with modern IDEs -- makes it unnecessarily hard for new or occasional users for getting started.
  8. unit-test driven development N.B. this to minimise errors, loop-holes, and to detect potential regression early-on as part of a general CI/CD strategy, but also to continuously re-evaluate design choices and quantitative evolution of the performance (for both: potential regressions and/or improvements, if possible).
  9. free- and open-source code basis w/o strings-attached:
    • it is important to us that this code can be re-used, built- and improved-upon by anybody and is not limited by unnecessary hurdles to due proprietary or IP-protected interfaces or licenses. N.B. we chose the LGPLv3 license in order that this remains free for future use, and to foster evolution of ideas and further developments that build upon this. See also this.

Some Serialiser Performance Comparison Results

The following examples are qualitative and primarily used to verify that our implementation is not significantly slower than another reference implementation and to document possible performance regression when refactoring the code base. Example output of SerialiserQuickBenchmark.java which compares the map-only, custom and full-pojo-to-pojo (de-)serialisation performance for the given low-level wire-format: 

Example output - numbers should be compared relatively (nIterations = 100000):
(openjdk 11.0.7 2020-04-14, ASCII-only, nSizePrimitiveArrays = 10, nSizeString = 100, nestedClassRecursion = 1)
[..] more string-heavy TestDataClass
- run 1
- JSON Serializer (Map only)  throughput = 371.4 MB/s for 5.2 kB per test run (took 1413.0 ms)
- CMW Serializer (Map only) throughput = 220.2 MB/s for 6.3 kB per test run (took 2871.0 ms)
- CmwLight Serializer (Map only)  throughput = 683.1 MB/s for 6.4 kB per test run (took 935.0 ms)
- IO Serializer (Map only)  throughput = 810.0 MB/s for 7.4 kB per test run (took 908.0 ms)

- FlatBuffers (custom FlexBuffers) throughput = 173.7 MB/s for 6.1 kB per test run (took 3536.0 ms)
- CmwLight Serializer (custom) throughput = 460.5 MB/s for 6.4 kB per test run (took 1387.0 ms)
- IO Serializer (custom) throughput = 545.0 MB/s for 7.3 kB per test run (took 1344.0 ms)

- JSON Serializer (POJO) throughput = 53.8 MB/s for 5.2 kB per test run (took 9747.0 ms)
- CMW Serializer (POJO) throughput = 182.8 MB/s for 6.3 kB per test run (took 3458.0 ms)
- CmwLight Serializer (POJO) throughput = 329.2 MB/s for 6.3 kB per test run (took 1906.0 ms)
- IO Serializer (POJO) throughput = 374.9 MB/s for 7.2 kB per test run (took 1925.0 ms)

[..] more primitive-array-heavy TestDataClass
(openjdk 11.0.7 2020-04-14, UTF8, nSizePrimitiveArrays = 1000, nSizeString = 0, nestedClassRecursion = 0)
- run 1
- JSON Serializer (Map only)  throughput = 350.7 MB/s for 34.3 kB per test run (took 9793.0 ms)
- CMW Serializer (Map only) throughput = 1.7 GB/s for 29.2 kB per test run (took 1755.0 ms)
- CmwLight Serializer (Map only)  throughput = 6.7 GB/s for 29.2 kB per test run (took 437.0 ms)
- IO Serializer (Map only)  throughput = 6.1 GB/s for 29.7 kB per test run (took 485.0 ms)

- FlatBuffers (custom FlexBuffers) throughput = 123.1 MB/s for 30.1 kB per test run (took 24467.0 ms)
- CmwLight Serializer (custom) throughput = 3.9 GB/s for 29.2 kB per test run (took 751.0 ms)
- IO Serializer (custom) throughput = 3.8 GB/s for 29.7 kB per test run (took 782.0 ms)

- JSON Serializer (POJO) throughput = 31.7 MB/s for 34.3 kB per test run (took 108415.0 ms)
- CMW Serializer (POJO) throughput = 1.5 GB/s for 29.2 kB per test run (took 1924.0 ms)
- CmwLight Serializer (POJO) throughput = 3.5 GB/s for 29.1 kB per test run (took 824.0 ms)
- IO Serializer (POJO) throughput = 3.4 GB/s for 29.7 kB per test run (took 870.0 ms)

A more thorough test using the Java micro-benchmark framework JMH output of SerialiserBenchmark.java for a string-heavy and for a numeric-data-heavy test data class: 

Benchmark                                     (testClassId)   Mode  Cnt       Score      Error  Units
SerialiserBenchmark.customCmwLight             string-heavy  thrpt   10   49954.479 ±  560.726  ops/s
SerialiserBenchmark.customCmwLight            numeric-heavy  thrpt   10   22433.828 ±  195.939  ops/s
SerialiserBenchmark.customFlatBuffer           string-heavy  thrpt   10   18446.085 ±   71.311  ops/s
SerialiserBenchmark.customFlatBuffer          numeric-heavy  thrpt   10     233.869 ±    7.314  ops/s
SerialiserBenchmark.customIoSerialiser         string-heavy  thrpt   10   53638.035 ±  367.122  ops/s
SerialiserBenchmark.customIoSerialiser        numeric-heavy  thrpt   10   24277.732 ±  200.380  ops/s
SerialiserBenchmark.customIoSerialiserOptim    string-heavy  thrpt   10   79759.984 ±  799.944  ops/s
SerialiserBenchmark.customIoSerialiserOptim   numeric-heavy  thrpt   10   24192.169 ±  419.019  ops/s
SerialiserBenchmark.customJson                 string-heavy  thrpt   10   17619.026 ±  250.917  ops/s
SerialiserBenchmark.customJson                numeric-heavy  thrpt   10     138.461 ±    2.972  ops/s
SerialiserBenchmark.mapCmwLight                string-heavy  thrpt   10   79273.547 ± 2487.931  ops/s
SerialiserBenchmark.mapCmwLight               numeric-heavy  thrpt   10   67374.131 ±  954.149  ops/s
SerialiserBenchmark.mapIoSerialiser            string-heavy  thrpt   10   81295.197 ± 2391.616  ops/s
SerialiserBenchmark.mapIoSerialiser           numeric-heavy  thrpt   10   67701.564 ± 1062.641  ops/s
SerialiserBenchmark.mapIoSerialiserOptimized   string-heavy  thrpt   10  115008.285 ± 2390.426  ops/s
SerialiserBenchmark.mapIoSerialiserOptimized  numeric-heavy  thrpt   10   68879.735 ± 1403.197  ops/s
SerialiserBenchmark.mapJson                    string-heavy  thrpt   10   14474.142 ± 1227.165  ops/s
SerialiserBenchmark.mapJson                   numeric-heavy  thrpt   10     163.928 ±    0.968  ops/s
SerialiserBenchmark.pojoCmwLight               string-heavy  thrpt   10   41821.232 ±  217.594  ops/s
SerialiserBenchmark.pojoCmwLight              numeric-heavy  thrpt   10   33820.451 ±  568.264  ops/s
SerialiserBenchmark.pojoIoSerialiser           string-heavy  thrpt   10   41899.128 ±  940.030  ops/s
SerialiserBenchmark.pojoIoSerialiser          numeric-heavy  thrpt   10   33918.815 ±  376.551  ops/s
SerialiserBenchmark.pojoIoSerialiserOptim      string-heavy  thrpt   10   53811.486 ±  920.474  ops/s
SerialiserBenchmark.pojoIoSerialiserOptim     numeric-heavy  thrpt   10   32463.267 ±  635.326  ops/s
SerialiserBenchmark.pojoJson                   string-heavy  thrpt   10   23327.701 ±  288.871  ops/s
SerialiserBenchmark.pojoJson                  numeric-heavy  thrpt   10     161.396 ±    3.040  ops/s
SerialiserBenchmark.pojoJsonCodeGen            string-heavy  thrpt   10   23586.818 ±  470.233  ops/s
SerialiserBenchmark.pojoJsonCodeGen           numeric-heavy  thrpt   10     163.250 ±    1.254  ops/s

N.B. The 'FlatBuffer' implementation is bit of an outlier and uses internally FlatBuffer's FlexBuffer builder which does not support and/or is not optimised for large primitive arrays. FlexBuffer was chosen primarily for comparison since it supported flexible compile/run-time map-type structures similar to the other implementations, whereas the faster Protobuf and Flatbuffer builder require IDL-based desciptions that are used during compile-time to generate the necessary data-serialiser stubs.

JSON-compatible strings are easy to construct and write. Nevertheless, we chose the Json-Itererator library as backend for implementing the JsonSerialiser for purely pragmatic reasons and to initially avoid common pitfalls in implementing a robust JSON deserialiser. The JsonSelectionBenchmark.java compares the choice with several other commonly used JSON serialisation libraries for a string-heavy and a numeric-data-heavy test data class:

Benchmark                                   (testClassId)   Mode  Cnt      Score     Error  Units
 JsonSelectionBenchmark.pojoFastJson          string-heavy  thrpt   10  12857.850 ± 109.050  ops/s
 JsonSelectionBenchmark.pojoFastJson         numeric-heavy  thrpt   10     91.458 ±   0.437  ops/s
 JsonSelectionBenchmark.pojoGson              string-heavy  thrpt   10   6253.698 ±  50.267  ops/s
 JsonSelectionBenchmark.pojoGson             numeric-heavy  thrpt   10     48.215 ±   0.265  ops/s
 JsonSelectionBenchmark.pojoJackson           string-heavy  thrpt   10  16563.604 ± 244.329  ops/s
 JsonSelectionBenchmark.pojoJackson          numeric-heavy  thrpt   10    135.780 ±   1.074  ops/s
 JsonSelectionBenchmark.pojoJsonIter          string-heavy  thrpt   10  10733.539 ±  35.605  ops/s
 JsonSelectionBenchmark.pojoJsonIter         numeric-heavy  thrpt   10     86.629 ±   1.122  ops/s
 JsonSelectionBenchmark.pojoJsonIterCodeGen   string-heavy  thrpt   10  41048.034 ± 396.628  ops/s
 JsonSelectionBenchmark.pojoJsonIterCodeGen  numeric-heavy  thrpt   10    377.412 ±   9.755  ops/s

Performance was not of primary concern for us, since JSON-based protocols are anyway slow. The heavy penalty for numeric-heavy data is largely related to the inefficient string representation of double values.

References