Benchmarking

In this benchmark we analyze the various way design the .proto message to encode an list of 32 bit integer:

Default

Using the OCaml int32 and list type:

repeated int32 l = 1;

ocaml_type = int_t extension

Using the int and list OCaml type:

repeated int32 l = 1 [(ocaml_type)=int_t];

ocaml_type = int_t / ocaml_container = repeated_field extensions

Using int type and the Runtime container Pbrt.Repeated_field:

repeated int32 l = 1 [(ocaml_type) = int_t, (ocaml_container) = repeated_field];

Packed encoding and all ocaml extensions

repeated int32 l = 1 [(ocaml_type) = int_t, (ocaml_container) = repeated_field];

Results in seconds for Decoding

Environment configuration:

OS: OSX Version 10.9.5
Processor: 1.4 Ghz Intel Core i5
Memory: 4 GB 1600 MHz DDR3

Results:

Size:	int32_list	int_list	int_repeated	int_packed_repeated
100	0.00001	0.00001	0.00001	0.00001
1000	0.00059	0.00008	0.00016	0.00006
50000	0.01324	0.00995	0.00593	0.00375
1000000	0.34267	0.23039	0.08685	0.05702
10000000	3.47812	2.35051	0.81929	0.67300

Conclusion

The results above are completely expected and can be explained:

Using int is more efficient than int32 in OCaml. This is because int32 is boxed.
Using the Runtime Pbrt.Repeated_field container improves performance over list. This data imperative data structure is optimzed for push back operation.
Using packed encoding improve performance: As explained by the protobuffer documentation, this encoding is more efficient.

The program benchs/bench.ml can be used to run a suite of benchmarks. For example, to run them all:

$ dune exec --profile=release benchs/benchs.exe -- --all

or to run only the varint encoding benchmarks:

dune exec --profile=release benchs/benchs.exe -- -p varint.enc

Provide feedback