We should consider switching our JSON parsing to: GitHub - minio/simdjson-go: Golang port of simdjson: parsing gigabytes of JSON per second
It claims to be 10x faster than Go’s JSON parser.
@mrjn May be you can consider json-iterator(GitHub - json-iterator/go: A high-performance 100% compatible drop-in replacement of "encoding/json") which is more mature and also 10x faster, and it 100% compatibility with standard Go’s JSON lib.
2 Likes
Is simdjson faster without a GPU (are the other SIMD implementations? I thought it’s currently only available to GPUs and very specialized hardware)?
We should probably run a poll to figure out how many of our deployments run on machines that have access to some SIMD (they have a handy function to check if it’s compatible).
json-iterator/go doesn’t seem to have any such requirements.
SIMD has been around for a while now. I think most modern intel chips support it. FWIW, typically these implementations do also have a fail-back mode, where they can switch to a normal Go implementation on platforms which don’t support SSE.
Though the json-iterator approach looks better, it has a lot more stars too, so more popular in the community.
Pretty much all cpus nowadays have some sort of SIMD architecture, so this shouldn’t be a concern.
I’d love to see benchmarks on real dgraph data!
I’ve found that simdjson-go has a nice benchmarks_test.go which compares it to std-json and jsoniter. A summary of this can be found here: GitHub - minio/simdjson-go: Golang port of simdjson: parsing gigabytes of JSON per second
BenchmarkApache_builds-12 8571 133707 ns/op 951.90 MB/s 1001 B/op 21 allocs/op
BenchmarkEncodingJsonApache_builds-12 745 1547654 ns/op 82.24 MB/s 464078 B/op 9717 allocs/op
BenchmarkJsoniterApache_builds-12 1131 1036249 ns/op 122.82 MB/s 520837 B/op 13249 allocs/op
BenchmarkCanada-12 31 34580642 ns/op 65.10 MB/s 12463564 B/op 333536 allocs/op
BenchmarkEncodingJsonCanada-12 25 46280032 ns/op 48.64 MB/s 12260723 B/op 392536 allocs/op
BenchmarkJsoniterCanada-12 16 65812488 ns/op 34.20 MB/s 16636987 B/op 665989 allocs/op
BenchmarkCitm_catalog-12 560 2155356 ns/op 801.35 MB/s 223779 B/op 14501 allocs/op
BenchmarkEncodingJsonCitm_catalog-12 56 21589300 ns/op 80.00 MB/s 5218456 B/op 95480 allocs/op
BenchmarkJsoniterCitm_catalog-12 99 11787867 ns/op 146.52 MB/s 5663367 B/op 118756 allocs/op
BenchmarkGithub_events-12 13824 86082 ns/op 756.63 MB/s 1614 B/op 119 allocs/op
BenchmarkEncodingJsonGithub_events-12 1642 720462 ns/op 90.40 MB/s 187511 B/op 3331 allocs/op
BenchmarkJsoniterGithub_events-12 2608 459739 ns/op 141.67 MB/s 222902 B/op 4443 allocs/op
BenchmarkGsoc_2018-12 838 1426016 ns/op 2333.66 MB/s 13275 B/op 66 allocs/op
BenchmarkEncodingJsonGsoc_2018-12 46 25282600 ns/op 131.63 MB/s 7064482 B/op 58616 allocs/op
BenchmarkJsoniterGsoc_2018-12 66 18439356 ns/op 180.47 MB/s 11802709 B/op 90914 allocs/op
BenchmarkInstruments-12 2926 397129 ns/op 554.85 MB/s 6398 B/op 1792 allocs/op
BenchmarkEncodingJsonInstruments-12 367 3234329 ns/op 68.13 MB/s 889352 B/op 13337 allocs/op
BenchmarkJsoniterInstruments-12 554 2162449 ns/op 101.90 MB/s 998816 B/op 18777 allocs/op
BenchmarkMarine_ik-12 30 33599490 ns/op 88.79 MB/s 8656960 B/op 436577 allocs/op
BenchmarkEncodingJsonMarine_ik-12 16 66812538 ns/op 44.65 MB/s 22791466 B/op 614777 allocs/op
BenchmarkJsoniterMarine_ik-12 19 59789532 ns/op 49.90 MB/s 25483422 B/op 692512 allocs/op
BenchmarkMesh-12 100 10199999 ns/op 70.94 MB/s 2694608 B/op 134240 allocs/op
BenchmarkEncodingJsonMesh-12 79 15341778 ns/op 47.17 MB/s 7161936 B/op 149505 allocs/op
BenchmarkJsoniterMesh-12 66 17712080 ns/op 40.85 MB/s 8020255 B/op 184138 allocs/op
BenchmarkMesh_pretty-12 99 11808081 ns/op 133.58 MB/s 3312611 B/op 134240 allocs/op
BenchmarkEncodingJsonMesh_pretty-12 54 22203720 ns/op 71.04 MB/s 7477376 B/op 149505 allocs/op
BenchmarkJsoniterMesh_pretty-12 54 21444441 ns/op 73.56 MB/s 8657071 B/op 204038 allocs/op
BenchmarkNumbers-12 525 2276186 ns/op 65.95 MB/s 802246 B/op 30030 allocs/op
BenchmarkEncodingJsonNumbers-12 487 2515404 ns/op 59.68 MB/s 1066432 B/op 20026 allocs/op
BenchmarkJsoniterNumbers-12 354 3347443 ns/op 44.85 MB/s 1226480 B/op 30038 allocs/op
BenchmarkRandom-12 1052 1134981 ns/op 449.77 MB/s 10323 B/op 2067 allocs/op
BenchmarkEncodingJsonRandom-12 124 9548385 ns/op 53.46 MB/s 2994630 B/op 70084 allocs/op
BenchmarkJsoniterRandom-12 169 7147923 ns/op 71.42 MB/s 3197255 B/op 88091 allocs/op
BenchmarkTwitter-12 1642 735079 ns/op 859.11 MB/s 15740 B/op 1565 allocs/op
BenchmarkEncodingJsonTwitter-12 148 8047331 ns/op 78.47 MB/s 2264518 B/op 31759 allocs/op
BenchmarkJsoniterTwitter-12 220 5409089 ns/op 116.75 MB/s 2461476 B/op 45043 allocs/op
BenchmarkTwitterescaped-12 1016 1176180 ns/op 478.17 MB/s 16347 B/op 1565 allocs/op
BenchmarkEncodingJsonTwitterescaped-12 152 7888163 ns/op 71.30 MB/s 2366189 B/op 31757 allocs/op
BenchmarkJsoniterTwitterescaped-12 182 6516472 ns/op 86.31 MB/s 2640968 B/op 47198 allocs/op
BenchmarkUpdate_center-12 1689 712848 ns/op 747.95 MB/s 2594 B/op 57 allocs/op
BenchmarkEncodingJsonUpdate_center-12 153 7869284 ns/op 67.75 MB/s 2768037 B/op 49089 allocs/op
BenchmarkJsoniterUpdate_center-12 190 6242113 ns/op 85.42 MB/s 3090373 B/op 66758 allocs/op
I can see jsoniter being very similar to std-json, while simdjson is way faster.
Benchmarks on jsoniter are a few years old, and it seems std-json has caught up. I really like this article: Go: Is the encoding/json Package Really Slow? | by Vincent Blanchon | A Journey With Go | Medium which compares std-json performance improvements between different go versions.