So, it is possible (by changing slightly the interface of encoders to
avoid an intermediate structure), to obtain the following (I also added
a benchmark for decoding, comparing cconv to ppx_deriving_yojson):

====== BEGIN BENCH ======

% ./run_bench.native
encoding...

benchmark points
Throughputs for "manual", "cconv", "deriving_yojson" each running for at least 4 CPU seconds:
         manual:  4.18 WALL ( 4.18 usr +  0.00 sys =  4.18 CPU) @ 2826676.15/s (n=11818333)
          cconv:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 941618.68/s (n=3951032)
deriving_yojson:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 2824134.70/s (n=11867014)
                     Rate           cconv deriving_yojson          manual
          cconv  941619/s              --            -67%            -67%
deriving_yojson 2824135/s            200%              --             -0%
         manual 2826676/s            200%              0%              --

benchmark terms
Throughputs for "manual", "cconv", "deriving_yojson" each running for at least 4 CPU seconds:
         manual:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 1125111.48/s (n=4723218)
          cconv:  4.21 WALL ( 4.21 usr +  0.00 sys =  4.21 CPU) @ 789967.68/s (n=3324184)
deriving_yojson:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 1100920.75/s (n=4626069)
                     Rate           cconv deriving_yojson          manual
          cconv  789968/s              --            -28%            -30%
deriving_yojson 1100921/s             39%              --             -2%
         manual 1125111/s             42%              2%              --


decoding...

benchmark points
Throughputs for "cconv", "deriving_yojson" each running for at least 3 CPU seconds:
          cconv:  3.16 WALL ( 3.16 usr +  0.00 sys =  3.16 CPU) @ 493501.11/s (n=1558970)
deriving_yojson:  3.15 WALL ( 3.15 usr +  0.00 sys =  3.15 CPU) @ 1248812.96/s (n=3932512)
                     Rate           cconv deriving_yojson
          cconv  493501/s              --            -60%
deriving_yojson 1248813/s            153%              --

benchmark terms
Throughputs for "cconv", "deriving_yojson" each running for at least 3 CPU seconds:
          cconv:  3.12 WALL ( 3.12 usr +  0.00 sys =  3.12 CPU) @ 577372.88/s (n=1800826)
deriving_yojson:  3.14 WALL ( 3.14 usr +  0.00 sys =  3.14 CPU) @ 1492303.95/s (n=4688819)
                     Rate           cconv deriving_yojson
          cconv  577373/s              --            -61%
deriving_yojson 1492304/s            158%              --
./run_bench.native  45.01s user 1.16s system 99% cpu 46.181 total

====== END BENCH ======

It is still slower to encode records (intermediate list), and decoding also
has some overhead. However, this is the cost of translating
between the type and the JSON representation; I think it should be
negligible compared to the actual IO + printing/parsing cost.

The new, more efficient interface will probably appear in a future release.
With ppx_deriving_cconv that shouldn't be too big a problem...

Cheers,

Le Mon, 01 Dec 2014, Simon Cruanes a écrit :
> Le Mon, 01 Dec 2014, Thomas Gazagnaire a écrit :
> > Do you have any benchmarks to compare CConv and similar camlp4 generators?
> 
> Hi Thomas,
> 
> I hadn't, but I just wrote very basic ones to compare with
> ppx_deriving_yojson (should be similar to camlp4). The code is at
> https://github.com/c-cube/cconv/blob/e80ab0e6c458a01b419ea69c7f41d0a350aebbad/bench/run_bench.ml
> 
> It only compares times for encoding into Json right now, with the
> following results (recursive records first, recursive terms then;
> "manual" is a handwritten encoding function, "cconv" the combinators
> version, and "deriving_yojson" uses @whitequark's nice deriver):
> 
> % ./run_bench.native
> 
> benchmark points
> Throughputs for "manual", "cconv", "deriving_yojson" each running for at least 4 CPU seconds:
>          manual:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 3057270.82/s (n=12846652)
>           cconv:  4.21 WALL ( 4.21 usr +  0.00 sys =  4.21 CPU) @ 784724.92/s (n=3300553)
> deriving_yojson:  4.21 WALL ( 4.21 usr +  0.00 sys =  4.21 CPU) @ 3065779.07/s (n=12891601)
>                      Rate           cconv          manual deriving_yojson
>           cconv  784725/s              --            -74%            -74%
>          manual 3057271/s            290%              --             -0%
> deriving_yojson 3065779/s            291%              0%              --
> 
> benchmark terms
> Throughputs for "manual", "cconv", "deriving_yojson" each running for at least 4 CPU seconds:
>          manual:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 1679609.71/s (n=7057720)
>           cconv:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 726619.43/s (n=3051075)
> deriving_yojson:  4.20 WALL ( 4.20 usr +  0.00 sys =  4.20 CPU) @ 1624740.65/s (n=6822286)
>                      Rate           cconv deriving_yojson          manual
>           cconv  726619/s              --            -55%            -57%
> deriving_yojson 1624741/s            124%              --             -3%
>          manual 1679610/s            131%              3%              --


-- 
Simon

http://weusepgp.info/
key 49AA62B6, fingerprint 949F EB87 8F06 59C6 D7D3  7D8D 4AC0 1D08 49AA 62B6