Awesome

simdjsone

An implementation of the fastest JSON parser for Erlang/Elixir using the C++ simdjson NIF library. The decoding speed of this parser is about 2.5 times faster than jiffy.

See full documentation of the C++ library.

Only a subset of functionality is implemented:

Ability to decode JSON terms using a main scheduler or a dirty scheduler based on the size of JSON binary input.
Ability to cache the decoded term, and access its key/value pairs using a NIF based simdjson:get/2 function.
The resources stored in the cache will get automatically garbage collected when the owner process dies.

For small JSON objects simdjsone is about twice faster than jiffy and for large JSON objects, it's about 30% faster than jiffy.

Decoding

The following decoding options are supported in decode(String, Options):

return_maps - decode JSON object as map (this is default)
object_as_tuple - decode JSON object as a proplist wrapped in a tuple
dedup_keys - eliminate duplicate keys from a JSON object
use_nil - decode JSON "null" as nil
{null_term, V} - use the given value V for a JSON "null"

Encoding

The following encoding options are supported in encode(String, Options):

uescape - escape UTF-8 sequences to produce a 7-bit clean output
pretty - return JSON using two-space indentation
use_nil - encode the atom nil as null
escape_fwd_slash - escape the / character (useful when encoding URLs)
{bytes_per_red, N} - where N >= 0 - This controls the number of bytes that the NIF library will process as an equivalent to a reduction. Each 20 reductions we consume 1% of our allocated time slice for the current process. When the Erlang VM indicates we need to return from the NIF.

NOTE: Since the simdjson library currently doen't have an implementation of a JSON encoder, the encoding implementation is the jiffy's modified encoder optimized for speed when encoding integers.

The implementation includes simdjson:int_to_bin/1 function that is about 30% faster than erlang:integer_to_binary/1, but it's limited to integers in range: (-1 bsl 63) <= I <= (1 bsl 62).

Author

Serge Aleynikov

Installation

Include the following dependency in your project.

Erlang (rebar.config):

{deps, [{simdjsone, "0.1"}]}.

Elixir (mix.exs):

def deps() do
  [{:simdjsone, "~> 0.2"}]
end

Build

Erlang:

$ make deps compile

Elixir

$ MIX_ENV=dev make deps compile

Simple JSON decoding

1> simdjson:decode("{\"a\": [1,2,3], \"b\": 123, \"c\": 12.234}").
#{<<"a">> => [1,2,3],<<"b">> => 123,<<"c">> => 12.234}

Cached JSON decoding

After calling the simdjson:parse/1, the function simdjson:get/2 returns the value stored in a given path:

1> Ref = simdjson:parse("{\"a\": [1,2,3], \"b\": 123, \"c\": 12.234}").
#Ref<0.1852532992.2458255361.217635>
2> simdjson:get(Ref, "/a").
[1,2,3]
3> simdjson:get(Ref, "/b").
123
4> simdjson:get(Ref, "/a/c").
12.234
ok

JSON encoding

1> simdjson:encode(#{a => [1,2,3], <<"b">> => 123, c => 12.234}).
<<"{\"b\":123,\"a\":[1,2,3],\"c\":12.234}">>
2> simdjson:encode({[{a, [1,2,3]}, {<<"b">>, 123}, {c, 12.234}]}).
<<"{\"a\":[1,2,3],\"b\":123,\"c\":12.234}">>

Performance Benchmark

To run the performance benchmark of simdjsone against jiffy and thoas do the following (prefix the command with CXX=clang++ for using Clang C++ compiler):

$ make benchmark
=== Benchmark (file size: 616.7K) ===
   simdjsone:   2526.420us
      euneus:   5314.820us
       thoas:   5452.380us
        json:   5541.130us
       jiffy:  10182.650us

=== Benchmark (file size: 1.3K) ===
   simdjsone:      6.940us
       thoas:      8.270us
      euneus:      9.790us
        json:     10.000us
       jiffy:     16.640us

=== Benchmark (file size: 0.1K) ===
   simdjsone:      1.930us
      euneus:      2.300us
        json:      2.500us
       jiffy:      2.770us
       thoas:      2.830us

If you have Elixir installed, the project also includes a benchmark for the jason and poison Elixir parsers. The Elixir benchmarks are more exhaustive:

$ MIX_ENV=test make benchmark
=== Benchmark (file size: 616.7K) ===

Name                ips        average  deviation         median         99th %
simdjsone        344.00        2.91 ms    ±18.13%        2.63 ms        4.82 ms
poison           154.88        6.46 ms    ±10.31%        6.25 ms        9.96 ms
jason            153.87        6.50 ms    ±11.37%        6.30 ms       10.18 ms
thaos            147.71        6.77 ms    ±11.53%        6.52 ms       10.64 ms
euneus           142.47        7.02 ms    ±22.71%        6.50 ms       13.47 ms
jiffy             78.74       12.70 ms    ±12.83%       12.52 ms       21.65 ms

Comparison:
simdjsone        344.00
poison           154.88 - 2.22x slower +3.55 ms
jason            153.87 - 2.24x slower +3.59 ms
thaos            147.71 - 2.33x slower +3.86 ms
euneus           142.47 - 2.41x slower +4.11 ms
jiffy             78.74 - 4.37x slower +9.79 ms

Memory usage statistics:

Name         Memory usage
simdjsone      0.00153 MB
poison            1.84 MB - 1200.09x memory usage +1.84 MB
jason             1.81 MB - 1182.38x memory usage +1.81 MB
thaos             1.81 MB - 1182.31x memory usage +1.81 MB
euneus            1.87 MB - 1219.47x memory usage +1.87 MB
jiffy             3.19 MB - 2077.98x memory usage +3.19 MB

**All measurements for memory usage were the same**

=== Benchmark (file size: 1.3K) ===

Name                ips        average  deviation         median         99th %
simdjsone      177.77 K        5.63 μs   ±187.84%        4.90 μs       16.90 μs
euneus         101.90 K        9.81 μs    ±94.81%           9 μs       22.30 μs
poison          98.29 K       10.17 μs    ±87.43%        9.40 μs       23.30 μs
jason           95.18 K       10.51 μs    ±99.89%        9.50 μs       30.70 μs
jiffy           88.49 K       11.30 μs   ±259.91%        9.30 μs       32.60 μs
thaos           84.36 K       11.85 μs    ±57.49%       10.90 μs       25.90 μs

Comparison:
simdjsone      177.77 K
euneus         101.90 K - 1.74x slower +4.19 μs
poison          98.29 K - 1.81x slower +4.55 μs
jason           95.18 K - 1.87x slower +4.88 μs
jiffy           88.49 K - 2.01x slower +5.68 μs
thaos           84.36 K - 2.11x slower +6.23 μs

Memory usage statistics:

Name         Memory usage
simdjsone         1.57 KB
euneus            5.22 KB - 3.32x memory usage +3.65 KB
poison            5.57 KB - 3.55x memory usage +4 KB
jason             5.29 KB - 3.37x memory usage +3.72 KB
jiffy             1.55 KB - 0.99x memory usage -0.01563 KB
thaos             5.22 KB - 3.32x memory usage +3.65 KB

**All measurements for memory usage were the same**

=== Benchmark (file size: 0.1K) ===

Name                ips        average  deviation         median         99th %
simdjsone      820.20 K        1.22 μs  ±1445.57%        1.10 μs        3.30 μs
poison         576.53 K        1.73 μs  ±1017.41%        1.40 μs        4.00 μs
thaos          489.07 K        2.04 μs  ±1003.95%        1.80 μs        4.70 μs
euneus         483.75 K        2.07 μs   ±624.18%        1.80 μs        4.70 μs
jason          435.99 K        2.29 μs   ±752.86%           2 μs        5.70 μs
jiffy          311.21 K        3.21 μs   ±652.57%        2.50 μs        9.10 μs

Comparison:
simdjsone      820.20 K
poison         576.53 K - 1.42x slower +0.52 μs
thaos          489.07 K - 1.68x slower +0.83 μs
euneus         483.75 K - 1.70x slower +0.85 μs
jason          435.99 K - 1.88x slower +1.07 μs
jiffy          311.21 K - 2.64x slower +1.99 μs

Memory usage statistics:

Name         Memory usage
simdjsone         0.50 KB
poison            1.32 KB - 2.64x memory usage +0.82 KB
thaos             1.20 KB - 2.41x memory usage +0.70 KB
euneus            1.20 KB - 2.41x memory usage +0.70 KB
jason             1.27 KB - 2.55x memory usage +0.77 KB
jiffy             1.46 KB - 2.92x memory usage +0.96 KB

**All measurements for memory usage were the same**

TODO:

Add support for iterator
Add support for iterate_many and parse_many