See also: https://github.com/gz-c/skycoin-serialization-benchmarks - Benchmarks of different encoders on a coin.SignedBlock
Current benchmark status, after merging of skycoin/skycoin#1915
# go test -bench=. -run=XXX -benchmem=true ./src/daemon
benchmark old ns/op new ns/op delta
BenchmarkDeserializeRawGetPeersMessage-8 155 78.3 -49.48%
BenchmarkSerializeGetPeersMessage-8 139 140 +0.72%
BenchmarkDeserializeRawGivePeersMessage-8 2790 810 -70.97%
BenchmarkSerializeGivePeersMessage-8 2121 1360 -35.88%
BenchmarkDeserializeRawIntroductionMessage-8 1548 615 -60.27%
BenchmarkSerializeIntroductionMessage-8 2886 1237 -57.14%
BenchmarkDeserializeRawGiveBlocksMessage-8 235440 39867 -83.07%
BenchmarkSerializeGiveBlocksMessage-8 216667 52153 -75.93%
BenchmarkDeserializeRawAnnounceTxnsMessage-8 4443 662 -85.10%
BenchmarkSerializeAnnounceTxnsMessage-8 4104 766 -81.34%
BenchmarkDeserializeRawGiveTxnsMessage-8 67512 11008 -83.69%
BenchmarkSerializeGiveTxnsMessage-8 61087 14769 -75.82%
benchmark old allocs new allocs delta
BenchmarkDeserializeRawGetPeersMessage-8 2 1 -50.00%
BenchmarkSerializeGetPeersMessage-8 2 2 +0.00%
BenchmarkDeserializeRawGivePeersMessage-8 41 11 -73.17%
BenchmarkSerializeGivePeersMessage-8 31 17 -45.16%
BenchmarkDeserializeRawIntroductionMessage-8 22 7 -68.18%
BenchmarkSerializeIntroductionMessage-8 21 13 -38.10%
BenchmarkDeserializeRawGiveBlocksMessage-8 1277 392 -69.30%
BenchmarkSerializeGiveBlocksMessage-8 907 557 -38.59%
BenchmarkDeserializeRawAnnounceTxnsMessage-8 17 8 -52.94%
BenchmarkSerializeAnnounceTxnsMessage-8 7 8 +14.29%
BenchmarkDeserializeRawGiveTxnsMessage-8 377 116 -69.23%
BenchmarkSerializeGiveTxnsMessage-8 259 161 -37.84%
benchmark old bytes new bytes delta
BenchmarkDeserializeRawGetPeersMessage-8 16 8 -50.00%
BenchmarkSerializeGetPeersMessage-8 16 16 +0.00%
BenchmarkDeserializeRawGivePeersMessage-8 689 160 -76.78%
BenchmarkSerializeGivePeersMessage-8 384 160 -58.33%
BenchmarkDeserializeRawIntroductionMessage-8 384 112 -70.83%
BenchmarkSerializeIntroductionMessage-8 320 160 -50.00%
BenchmarkDeserializeRawGiveBlocksMessage-8 51146 16584 -67.58%
BenchmarkSerializeGiveBlocksMessage-8 15728 11760 -25.23%
BenchmarkDeserializeRawAnnounceTxnsMessage-8 1127 352 -68.77%
BenchmarkSerializeAnnounceTxnsMessage-8 176 240 +36.36%
BenchmarkDeserializeRawGiveTxnsMessage-8 15463 4856 -68.60%
BenchmarkSerializeGiveTxnsMessage-8 4496 3408 -24.20%
# go test -bench=. -run=XXX -benchmem=true ./src/cipher/encoder
benchmark old ns/op new ns/op delta
BenchmarkDeserializeRaw-8 3065 657 -78.56%
BenchmarkSerialize-8 1069 790 -26.10%
benchmark old allocs new allocs delta
BenchmarkDeserializeRaw-8 31 11 -64.52%
BenchmarkSerialize-8 13 7 -46.15%
benchmark old bytes new bytes delta
BenchmarkDeserializeRaw-8 883 216 -75.54%
BenchmarkSerialize-8 208 112 -46.15%
This following benchmark of the binary encoder was done with https://github.com/gz-c/gosercomp. The benchmarks require a lot of setup for the various encoders, but is described in the README of that repo.
Results generated on Mid-2015 Macbook Pro base model, commit a8ed0bd51047396993bcb2e16a070826618db03f, before any optimizations were made:
BenchmarkMarshalByJson-8 2000000 688 ns/op 128 B/op 2 allocs/op
BenchmarkUnmarshalByJson-8 500000 2212 ns/op 216 B/op 8 allocs/op
BenchmarkMarshalByXml-8 500000 3726 ns/op 4800 B/op 11 allocs/op
BenchmarkUnmarshalByXml-8 100000 13710 ns/op 3027 B/op 73 allocs/op
BenchmarkMarshalByMsgp-8 20000000 122 ns/op 80 B/op 1 allocs/op
BenchmarkUnmarshalByMsgp-8 10000000 222 ns/op 32 B/op 5 allocs/op
BenchmarkMarshalByProtoBuf-8 10000000 233 ns/op 48 B/op 1 allocs/op
BenchmarkUnmarshalByProtoBuf-8 3000000 606 ns/op 160 B/op 10 allocs/op
BenchmarkMarshalByGogoProtoBuf-8 10000000 132 ns/op 48 B/op 1 allocs/op
BenchmarkUnmarshalByGogoProtoBuf-8 3000000 387 ns/op 144 B/op 8 allocs/op
BenchmarkMarshalByFlatBuffers-8 5000000 361 ns/op 16 B/op 1 allocs/op
BenchmarkUnmarshalByFlatBuffers-8 2000000000 0.32 ns/op 0 B/op 0 allocs/op
BenchmarkUnmarshalByFlatBuffers_withFields-8 10000000 126 ns/op 0 B/op 0 allocs/op
BenchmarkMarshalByThrift-8 3000000 446 ns/op 64 B/op 1 allocs/op
BenchmarkUnmarshalByThrift-8 1000000 1300 ns/op 656 B/op 11 allocs/op
BenchmarkMarshalByAvro-8 3000000 541 ns/op 48 B/op 6 allocs/op
BenchmarkUnmarshalByAvro-8 500000 3257 ns/op 1672 B/op 62 allocs/op
BenchmarkMarshalByGencode-8 30000000 44.6 ns/op 0 B/op 0 allocs/op
BenchmarkUnmarshalByGencode-8 10000000 121 ns/op 32 B/op 5 allocs/op
BenchmarkMarshalByUgorjiCodecAndCbor-8 2000000 727 ns/op 112 B/op 3 allocs/op
BenchmarkUnmarshalByUgorjiCodecAndCbor-8 2000000 673 ns/op 48 B/op 6 allocs/op
BenchmarkMarshalByUgorjiCodecAndMsgp-8 2000000 692 ns/op 112 B/op 3 allocs/op
BenchmarkUnmarshalByUgorjiCodecAndMsgp-8 2000000 722 ns/op 48 B/op 6 allocs/op
BenchmarkMarshalByUgorjiCodecAndBinc-8 2000000 807 ns/op 112 B/op 3 allocs/op
BenchmarkUnmarshalByUgorjiCodecAndBinc-8 1000000 1107 ns/op 824 B/op 10 allocs/op
BenchmarkMarshalByUgorjiCodecAndJson-8 2000000 1046 ns/op 112 B/op 3 allocs/op
BenchmarkUnmarshalByUgorjiCodecAndJson-8 2000000 813 ns/op 48 B/op 6 allocs/op
BenchmarkMarshalByEasyjson-8 5000000 337 ns/op 128 B/op 1 allocs/op
BenchmarkUnmarshalByEasyjson-8 3000000 535 ns/op 32 B/op 5 allocs/op
BenchmarkMarshalByFfjson-8 1000000 1064 ns/op 424 B/op 9 allocs/op
BenchmarkUnmarshalByFfjson-8 1000000 1535 ns/op 480 B/op 13 allocs/op
BenchmarkMarshalByJsoniter-8 2000000 659 ns/op 96 B/op 2 allocs/op
BenchmarkUnmarshalByJsoniter-8 2000000 647 ns/op 32 B/op 5 allocs/op
BenchmarkUnmarshalByGJSON-8 1000000 1885 ns/op 624 B/op 7 allocs/op
BenchmarkMarshalByGoMemdump-8 300000 5328 ns/op 1032 B/op 30 allocs/op
BenchmarkUnmarshalByGoMemdump-8 1000000 1534 ns/op 2400 B/op 12 allocs/op
BenchmarkMarshalBySky-8 1000000 1140 ns/op 208 B/op 13 allocs/op
BenchmarkUnmarshalBySky-8 500000 3175 ns/op 883 B/op 31 allocs/op
BenchmarkMarshalByColfer-8 50000000 33.6 ns/op 0 B/op 0 allocs/op
BenchmarkUnmarshalByColfer-8 5000000 212 ns/op 96 B/op 6 allocs/op
BenchmarkMarshalByZebrapack-8 10000000 240 ns/op 182 B/op 0 allocs/op
BenchmarkUnmarshalByZebrapack-8 5000000 303 ns/op 32 B/op 5 allocs/op
BenchmarkMarshalByGotiny-8 3000000 465 ns/op 184 B/op 6 allocs/op
BenchmarkUnmarshalByGotiny-8 5000000 261 ns/op 88 B/op 2 allocs/op
BenchmarkMarshalByHprose-8 3000000 462 ns/op 210 B/op 1 allocs/op
BenchmarkUnmarshalByHprose-8 2000000 643 ns/op 288 B/op 9 allocs/op
BenchmarkMarshalBySereal-8 1000000 2288 ns/op 792 B/op 22 allocs/op
BenchmarkUnmarshalBySereal-8 2000000 755 ns/op 80 B/op 6 allocs/op
BenchmarkMarshalByMsgpackV2-8 1000000 2124 ns/op 192 B/op 4 allocs/op
BenchmarkUnmarshalByMsgpackv2-8 1000000 1657 ns/op 232 B/op 11 allocs/op
BenchmarkMarshalMsgZColorGroup-8 20000000 81.4 ns/op 80 B/op 1 allocs/op
BenchmarkAppendMsgZColorGroup-8 30000000 44.7 ns/op 313.29 MB/s 0 B/op 0 allocs/op
BenchmarkUnmarshalZColorGroup-8 20000000 86.9 ns/op 161.02 MB/s 0 B/op 0 allocs/op
BenchmarkEncodeZColorGroup-8 20000000 75.8 ns/op 184.78 MB/s 16 B/op 1 allocs/op
BenchmarkDecodeZColorGroup-8 10000000 136 ns/op 102.57 MB/s 0 B/op 0 allocs/op
gosercomp_test.go:99: json: 65 bytes
gosercomp_test.go:102: xml: 137 bytes
gosercomp_test.go:105: msgp: 47 bytes
gosercomp_test.go:108: protobuf: 36 bytes
gosercomp_test.go:111: gogoprotobuf: 36 bytes
gosercomp_test.go:115: flatbuffers: 108 bytes
gosercomp_test.go:121: thrift: 63 bytes
gosercomp_test.go:135: avro: 32 bytes
gosercomp_test.go:144: gencode: 34 bytes
gosercomp_test.go:150: UgorjiCodec_Cbor: 47 bytes
gosercomp_test.go:156: UgorjiCodec_Msgp: 47 bytes
gosercomp_test.go:162: UgorjiCodec_Bin: 53 bytes
gosercomp_test.go:164: UgorjiCodec_Json: 91 bytes
gosercomp_test.go:167: easyjson: 65 bytes
gosercomp_test.go:170: ffjson: 65 bytes
gosercomp_test.go:173: jsoniter: 65 bytes
gosercomp_test.go:177: memdump: 200 bytes
gosercomp_test.go:180: colfer: 35 bytes
gosercomp_test.go:183: sky: 52 bytes
gosercomp_test.go:186: zebrapack: 48 bytes
gosercomp_test.go:189: gotiny: 32 bytes
gosercomp_test.go:194: hprose: 32 bytes
gosercomp_test.go:198: sereal: 76 bytes
gosercomp_test.go:201: msgpackv2: 47 bytes
The DB storage uses the Sky encoder for objects. Slow deserialization can impact queries and scans. However, after profiling some visor.Visor methods, it seems that the reflect-based Sky encoder has only a small contribution to performance. Therefore, optimizing it further will not show much improvement (unless the profiling data has been misread).