/simdjson

Parsing gigabytes of JSON per second

Primary LanguageC++Apache License 2.0Apache-2.0

simdjson : Parsing gigabytes of JSON per second

A C++ library to see how fast we can parse JSON with complete validation.

JSON documents are everywhere on the Internet. Servers spend a lot of time parsing these documents. We want to accelerate the parsing of JSON per se using commonly available SIMD instructions as much as possible while doing full validation (including character encoding).

Some performance results

We can use a quarter or fewer instructions than a state-of-the-art parser like RapidJSON, and half as many as sajson. To our knowledge, simdjson is the first fully-validating JSON parser to run at gigabytes per second on commodity processors.

On a Skylake processor, the parsing speeds (in GB/s) of various processors on the twitter.json file are as follows.

parser GB/s
simdjson 2.2
RapidJSON encoding-validation 0.51
RapidJSON encoding-validation, insitu 0.71
sajson (insitu, dynamic) 0.70
sajson (insitu, static) 0.97
dropbox 0.14
fastjson 0.26
gason 0.85
ultrajson 0.42
jsmn 0.28
cJSON 0.34

Requirements

  • We support platforms like Linux or macOS, as well as Windows through Visual Studio 2017 or later.
  • A processor with AVX2 (i.e., Intel processors starting with the Haswell microarchitecture released 2013, and processors from AMD starting with the Ryzen)
  • A recent C++ compiler (e.g., GNU GCC or LLVM CLANG or Visual Studio 2017), we assume C++17. GNU GCC 7 or better or LLVM's clang 6 or better.
  • Some benchmark scripts assume bash and other common utilities, but they are optional.

License

This code is made available under the Apache License 2.0.

Under Windows, we build some tools using the windows/dirent_portable.h file (which is outside our library code): it under the liberal (business-friendly) MIT license.

Code example

#include "simdjson/jsonparser.h"

/...

const char * filename = ... //

// use whatever means you want to get a string of your JSON document
std::string_view p = get_corpus(filename);
ParsedJson pj;
pj.allocateCapacity(p.size()); // allocate memory for parsing up to p.size() bytes
bool is_ok = json_parse(p, pj); // do the parsing, return false on error
// parsing is done!
// You can safely delete the string content
free((void*)p.data());
// the ParsedJson document can be used here
// js can be reused with other json_parse calls.

It is also possible to use a simpler API if you do not mind having the overhead of memory allocation with each new JSON document:

#include "simdjson/jsonparser.h"

/...

const char * filename = ... //
std::string_view p = get_corpus(filename);
ParsedJson pj = build_parsed_json(p); // do the parsing
// you no longer need p at this point, can do aligned_free((void*)p.data())
if( ! pj.isValid() ) {
    // something went wrong
}

Usage: easy single-header version

See the "singleheader" repository for a single header version. See the included file "amalgamation_demo.cpp" for usage. This requires no specific build system: just copy the files in your project in your include path. You can then include them quite simply:

#include <iostream>
#include "simdjson.h"
#include "simdjson.cpp"
int main(int argc, char *argv[]) {
  const char * filename = argv[1]; 
  std::string_view p = get_corpus(filename);
  ParsedJson pj = build_parsed_json(p); // do the parsing
  if( ! pj.isValid() ) {
    std::cout << "not valid" << std::endl;
  } else {
    std::cout << "valid" << std::endl;
  }
  return EXIT_SUCCESS;
}

Note: In some settings, it might be desirable to precompile simdjson.cpp instead of including it.

Usage (old-school Makefile on platforms like Linux or macOS)

Requirements: recent clang or gcc, and make. We recommend at least GNU GCC/G++ 7 or LLVM clang 6. A system like Linux or macOS is expected.

To test:

make
make test

To run benchmarks:

make parse
./parse jsonexamples/twitter.json

Under Linux, the parse command gives a detailed analysis of the performance counters.

To run comparative benchmarks (with other parsers):

make benchmark

Usage (CMake on platforms like Linux or macOS)

Requirements: We require a recent version of cmake. On macOS, the easiest way to install cmake might be to use brew and then type

brew install cmake

There is an equivalent brew on Linux which works the same way as well.

You need a recent compiler like clang or gcc. We recommend at least GNU GCC/G++ 7 or LLVM clang 6. For example, you can install a recent compiler with brew:

brew install gcc@8

Optional: You need to tell cmake which compiler you wish to use by setting the CC and CXX variables. Under bash, you can do so with commands such as export CC=gcc-7 and export CXX=g++-7.

Building: While in the project repository, do the following:

mkdir build
cd build
cmake ..
make
make test

CMake will build a library. By default, it builds a shared library (e.g., libsimdjson.so on Linux).

You can build a static library:

mkdir buildstatic
cd buildstatic
cmake -DSIMDJSON_BUILD_STATIC=ON ..
make 
make test

In some cases, you may want to specify your compiler, especially if the default compiler on your system is too old. You may proceed as follows:

brew install gcc@8
mkdir build
cd build
export CXX=g++-8 CC=gcc-8
cmake ..
make
make test

Usage (CMake on Windows using Visual Studio)

We are assuming that you have a common Windows PC with at least Visual Studio 2017, and an x64 processor with AVX2 support (2013 Haswell or later).

  • Grab the simdjson code from GitHub, e.g., by cloning it using GitHub Desktop.
  • Install CMake. When you install it, make sure to ask that cmake be made available from the command line. Please choose a recent version of cmake.
  • Create a subdirectory within simdjson, such as VisualStudio.
  • Using a shell, go to this newly created directory.
  • Type cmake -DCMAKE_GENERATOR_PLATFORM=x64 .. in the shell while in the VisualStudio repository. (Alternatively, if you want to build a DLL, you may use the command line cmake -DCMAKE_GENERATOR_PLATFORM=x64 -DSIMDJSON_BUILD_STATIC=OFF ...)
  • This last command created a Visual Studio solution file in the newly created directory (e.g., simdjson.sln). Open this file in Visual Studio. You should now be able to build the project and run the tests. For example, in the Solution Explorer window (available from the View menu), right-click ALL_BUILD and select Build. To test the code, still in the Solution Explorer window, select RUN_TESTS and select Build.

Tools

  • json2json mydoc.json parses the document, constructs a model and then dumps back the result to standard output.
  • json2json -d mydoc.json parses the document, constructs a model and then dumps model (as a tape) to standard output. The tape format is described in the accompanying file tape.md.
  • minify mydoc.json minifies the JSON document, outputting the result to standard output. Minifying means to remove the unneeded white space characters.

Scope

We provide a fast parser. It fully validates the input according to the various specifications. The parser builds a useful immutable (read-only) DOM (document-object model) which can be later accessed.

To simplify the engineering, we make some assumptions.

  • We support UTF-8 (and thus ASCII), nothing else (no Latin, no UTF-16). We do not believe that this is a genuine limitation in the sense that we do not think that there is any serious application that needs to process JSON data without an ASCII or UTF-8 encoding.
  • We store strings as NULL terminated C strings. Thus we implicitly assume that you do not include a NULL character within your string, which is allowed technically speaking if you escape it (\u0000).
  • We assume AVX2 support which is available in all recent mainstream x86 processors produced by AMD and Intel. No support for non-x86 processors is included though it can be done. We plan to support ARM processors (help is invited).
  • In cases of failure, we just report a failure without any indication as to the nature of the problem. (This can be easily improved without affecting performance.)
  • As allowed by the specification, we allow repeated keys within an object (other parsers like sajson do the same).
  • Performance is optimized for JSON documents spanning at least a few kilobytes up to many megabytes: the performance issues with having to parse many tiny JSON documents or one truly enormous JSON document are different.

We do not aim to provide a general-purpose JSON library. A library like RapidJSON offers much more than just parsing, it helps you generate JSON and offers various other convenient functions. We merely parse the document.

Features

  • The input string is unmodified. (Parsers like sajson and RapidJSON use the input string as a buffer.)
  • We parse integers and floating-point numbers as separate types which allows us to support large 64-bit integers in [-9223372036854775808,9223372036854775808), like a Java long or a C/C++ long long. Among the parsers that differentiate between integers and floating-point numbers, not all support 64-bit integers. (For example, sajson rejects JSON files with integers larger than or equal to 2147483648. RapidJSON will parse a file containing an overly long integer like 18446744073709551616 as a floating-point number.) When we cannot represent exactly an integer as a signed 64-bit value, we reject the JSON document.
  • We do full UTF-8 validation as part of the parsing. (Parsers like fastjson, gason and dropbox json11 do not do UTF-8 validation.)
  • We fully validate the numbers. (Parsers like gason and ultranjson will accept [0e+] as valid JSON.)
  • We validate string content for unescaped characters. (Parsers like fastjson and ultrajson accept unescaped line breaks and tabs in strings.)

Architecture

The parser works in three stages:

  • Stage 1. (Find marks) Identifies quickly structure elements, strings, and so forth. We validate UTF-8 encoding at that stage.
  • Stage 2. (Structure building) Involves constructing a "tree" of sort (materialized as a tape) to navigate through the data. Strings and numbers are parsed at this stage.

Navigating the parsed document

Here is a code sample to dump back the parsed JSON to a string:

    ParsedJson::iterator pjh(pj);
    if (!pjh.isOk()) {
      std::cerr << " Could not iterate parsed result. " << std::endl;
      return EXIT_FAILURE;
    }
    compute_dump(pj);
    //
    // where compute_dump is :

void compute_dump(ParsedJson::iterator &pjh) {
  if (pjh.is_object()) {
    std::cout << "{";
    if (pjh.down()) {
      pjh.print(std::cout); // must be a string
      std::cout << ":";
      pjh.next();
      compute_dump(pjh); // let us recurse
      while (pjh.next()) {
        std::cout << ",";
        pjh.print(std::cout);
        std::cout << ":";
        pjh.next();
        compute_dump(pjh); // let us recurse
      }
      pjh.up();
    }
    std::cout << "}";
  } else if (pjh.is_array()) {
    std::cout << "[";
    if (pjh.down()) {
      compute_dump(pjh); // let us recurse
      while (pjh.next()) {
        std::cout << ",";
        compute_dump(pjh); // let us recurse
      }
      pjh.up();
    }
    std::cout << "]";
  } else {
    pjh.print(std::cout); // just print the lone value
  }
}

The following function will find all user.id integers:

void simdjson_traverse(std::vector<int64_t> &answer, ParsedJson::iterator &i) {
  switch (i.get_type()) {
  case '{':
    if (i.down()) {
      do {
        bool founduser = equals(i.get_string(), "user");
        i.next(); // move to value
        if (i.is_object()) {
          if (founduser && i.move_to_key("id")) {
            if (i.is_integer()) {
              answer.push_back(i.get_integer());
            }
            i.up();
          }
          simdjson_traverse(answer, i);
        } else if (i.is_array()) {
          simdjson_traverse(answer, i);
        }
      } while (i.next());
      i.up();
    }
    break;
  case '[':
    if (i.down()) {
      do {
        if (i.is_object_or_array()) {
          simdjson_traverse(answer, i);
        }
      } while (i.next());
      i.up();
    }
    break;
  case 'l':
  case 'd':
  case 'n':
  case 't':
  case 'f':
  default:
    break;
  }
}

In-depth comparisons

If you want to see how a wide range of parsers validate a given JSON file:

make allparserscheckfile
./allparserscheckfile myfile.json

For performance comparisons:

make parsingcompetition
./parsingcompetition myfile.json

For broader comparisons:

make allparsingcompetition
./allparsingcompetition myfile.json

Various References

Inspiring links:

Validating UTF-8 takes no more than 0.7 cycles per byte:

Remarks on JSON parsing

  • The JSON spec defines what a JSON parser is:

A JSON parser transforms a JSON text into another representation. A JSON parser MUST accept all texts that conform to the JSON grammar. A JSON parser MAY accept non-JSON forms or extensions. An implementation may set limits on the size of texts that it accepts. An implementation may set limits on the maximum depth of nesting. An implementation may set limits on the range and precision of numbers. An implementation may set limits on the length and character contents of strings.

  • JSON is not JavaScript:

All JSON is Javascript but NOT all Javascript is JSON. So {property:1} is invalid because property does not have double quotes around it. {'property':1} is also invalid, because it's single quoted while the only thing that can placate the JSON specification is double quoting. JSON is even fussy enough that {"property":.1} is invalid too, because you should have of course written {"property":0.1}. Also, don't even think about having comments or semicolons, you guessed it: they're invalid. (credit:https://github.com/elzr/vim-json)

  • The structural characters are:

    begin-array     =  [ left square bracket
begin-object    =  { left curly bracket
end-array       =  ] right square bracket
end-object      =  } right curly bracket
name-separator  = : colon
value-separator = , comma

Pseudo-structural elements

A character is pseudo-structural if and only if:

  1. Not enclosed in quotes, AND
  2. Is a non-whitespace character, AND
  3. It's preceding chararacter is either: (a) a structural character, OR (b) whitespace.

This helps as we redefine some new characters as pseudo-structural such as the characters 1, 1, G, n in the following:

{ "foo" : 1.5, "bar" : 1.5 GEOFF_IS_A_DUMMY bla bla , "baz", null }

Academic References

  • T.MĂĽhlbauer, W.Rödiger, R.Seilbeck, A.Reiser, A.Kemper, and T.Neumann. Instant loading for main memory databases. PVLDB, 6(14):1702–1713, 2013. (SIMD-based CSV parsing)
  • Mytkowicz, Todd, Madanlal Musuvathi, and Wolfram Schulte. "Data-parallel finite-state machines." ACM SIGARCH Computer Architecture News. Vol. 42. No. 1. ACM, 2014.
  • Lu, Yifan, et al. "Tree structured data processing on GPUs." Cloud Computing, Data Science & Engineering-Confluence, 2017 7th International Conference on. IEEE, 2017.
  • Sidhu, Reetinder. "High throughput, tree automata based XML processing using FPGAs." Field-Programmable Technology (FPT), 2013 International Conference on. IEEE, 2013.
  • Dai, Zefu, Nick Ni, and Jianwen Zhu. "A 1 cycle-per-byte XML parsing accelerator." Proceedings of the 18th annual ACM/SIGDA international symposium on Field programmable gate arrays. ACM, 2010.
  • Lin, Dan, et al. "Parabix: Boosting the efficiency of text processing on commodity processors." High Performance Computer Architecture (HPCA), 2012 IEEE 18th International Symposium on. IEEE, 2012. http://parabix.costar.sfu.ca/export/1783/docs/HPCA2012/final_ieee/final.pdf
  • Deshmukh, V. M., and G. R. Bamnote. "An empirical evaluation of optimization parameters in XML parsing for performance enhancement." Computer, Communication and Control (IC4), 2015 International Conference on. IEEE, 2015.
  • Moussalli, Roger, et al. "Efficient XML Path Filtering Using GPUs." ADMS@ VLDB. 2011.
  • Jianliang, Ma, et al. "Parallel speculative dom-based XML parser." High Performance Computing and Communication & 2012 IEEE 9th International Conference on Embedded Software and Systems (HPCC-ICESS), 2012 IEEE 14th International Conference on. IEEE, 2012.
  • Li, Y., Katsipoulakis, N.R., Chandramouli, B., Goldstein, J. and Kossmann, D., 2017. Mison: a fast JSON parser for data analytics. Proceedings of the VLDB Endowment, 10(10), pp.1118-1129. http://www.vldb.org/pvldb/vol10/p1118-li.pdf
  • Cameron, Robert D., et al. "Parallel scanning with bitstream addition: An xml case study." European Conference on Parallel Processing. Springer, Berlin, Heidelberg, 2011.
  • Cameron, Robert D., Kenneth S. Herdy, and Dan Lin. "High performance XML parsing using parallel bit stream technology." Proceedings of the 2008 conference of the center for advanced studies on collaborative research: meeting of minds. ACM, 2008.
  • Shah, Bhavik, et al. "A data parallel algorithm for XML DOM parsing." International XML Database Symposium. Springer, Berlin, Heidelberg, 2009.
  • Cameron, Robert D., and Dan Lin. "Architectural support for SWAR text processing with parallel bit streams: the inductive doubling principle." ACM Sigplan Notices. Vol. 44. No. 3. ACM, 2009.
  • Amagasa, Toshiyuki, Mana Seino, and Hiroyuki Kitagawa. "Energy-Efficient XML Stream Processing through Element-Skipping Parsing." Database and Expert Systems Applications (DEXA), 2013 24th International Workshop on. IEEE, 2013.
  • Medforth, Nigel Woodland. "icXML: Accelerating Xerces-C 3.1. 1 using the Parabix Framework." (2013).
  • Zhang, Qiang Scott. Embedding Parallel Bit Stream Technology Into Expat. Diss. Simon Fraser University, 2010.
  • Cameron, Robert D., et al. "Fast Regular Expression Matching with Bit-parallel Data Streams."
  • Lin, Dan. Bits filter: a high-performance multiple string pattern matching algorithm for malware detection. Diss. School of Computing Science-Simon Fraser University, 2010.
  • Yang, Shiyang. Validation of XML Document Based on Parallel Bit Stream Technology. Diss. Applied Sciences: School of Computing Science, 2013.
  • N. Nakasato, "Implementation of a parallel tree method on a GPU", Journal of Computational Science, vol. 3, no. 3, pp. 132-141, 2012.