Compare various data serialization libraries for C++.
This project does not have any external library dependencies. All (boost, thrift etc.) needed libraries are downloaded and built automatically, but you need enough free disk space to build all components. To build this project you need a compiler that supports C++11 features. Project was tested with GCC 4.8.2 (Ubuntu 14.04).
$ git clone https://github.com/thekvs/cpp-serializers.git
$ mkdir /path/to/build-root/
$ cd /path/to/build-root/
$ cmake /path/to/cpp-serializers -DCMAKE_BUILD_TYPE=Release
$ make
- Test all serializers, run each serializer 100000 times:
$ ./benchmark 100000
- Test only protobuf serializer, run it 100000 times:
$ ./benchmark 100000 protobuf
- Test protobuf and cereal serializers only, run each of them 100000 times:
$ ./benchmark 100000 protobuf cereal
Following results were obtained running 1000000 serialize-deserialize operations 50 times and then averaging results on a typical desktop computer with Intel Core i5 processor running Ubuntu 14.04. Exact versions of libraries used are:
- thrift 0.9.2
- protobuf 2.6.0
- boost 1.59.0
- msgpack 0.5.9
- cereal 1.1.2
- avro 1.7.7
- capnproto 0.5.2
- flatbuffers 1.3.0
| serializer | object's size | avg. total time |
|---|---|---|
| thrift-binary | 17017 | 13763 |
| thrift-compact | 11597 | 27017 |
| protobuf | 12571 | 21034 |
| boost | 17470 | 22945 |
| msgpack | 11902 | 23560 |
| cereal | 17416 | 10688 |
| avro | 12288 | 31750 |
For capnproto and flatbuffers since they already store data in a "serialized" form and serialization basically means getting pointer to the internal storage, we measure full build/serialize/deserialize cycle. In the case of other libraries we measure serialize/deserialize cycle of the already built data structure.
| serializer | object's size | avg. total time |
|---|---|---|
| capnproto | 17768 | 4396 |
| flatbuffers | 17632 | 12494 |
Size measured in bytes, time measured in milliseconds.