Thank you for sharing this interesting information, Gabriel!  We
benchmarked the code exclusively with 4.01.0 and didn't know about the
performance boost in printf.

I just measured the performance with
examples/kinsol/serial/kinFerTron_dns.ml, which was one of the
examples that had the most pronounced effect.  As you suggest, the
numbers are a lot closer with 4.02.1.  The median wall-clock times of
10 runs were:

printf,  OCaml 4.02.1: 1.76[s]
print_*, OCaml 4.02.1: 1.62[s]
printf,  OCaml 4.01.0: 2.04[s]
print_*, OCaml 4.01.0: 1.70[s]

The overhead of using printf is about 8% in 4.02.1, as opposed to
about 20% in 4.01.0.  So in 4.02 the effect is noticeably smaller,
though not unmeasurable.  We should note this in the doc in a future
release.

FYI, the experiment can be reproduced as follows (in bash syntax),
using the attached patch (referred to as
/tmp/kinFerTron_dns_printf.diff below):

# Start at the root of the source tree.
opam switch 4.02.1; eval `opam config env`
./configure && make clean all
cd examples/kinsol/serial
# Measure without modification.
make PERF_DATA_POINTS=10 kinFerTron_dns.opt.perf
../../utils/crunchperf -m kinFerTron_dns.opt.perf >
no-printf-4.02.1-kinFerTron_dns.opt.perf
# Apply patch and measure again.
patch -p4 < /tmp/kinFerTron_dns_printf.diff
make PERF_DATA_POINTS=10 kinFerTron_dns.opt.perf
../../utils/crunchperf -m kinFerTron_dns.opt.perf >
printf-4.02.1-kinFerTron_dns.opt.perf
# Undo changes.
patch -p4 -R < /tmp/kinFerTron_dns_printf.diff

cd ../../..
# Back at the root of the source tree.
opam switch 4.01.0; eval `opam config env`
# (Basically the same thing as above)
./configure && make clean all
cd examples/kinsol/serial
make PERF_DATA_POINTS=10 kinFerTron_dns.opt.perf
../../utils/crunchperf -m kinFerTron_dns.opt.perf >
no-printf-4.01.0-kinFerTron_dns.opt.perf
patch -p4 < /tmp/kinFerTron_dns_printf.diff
make PERF_DATA_POINTS=10 kinFerTron_dns.opt.perf
../../utils/crunchperf -m kinFerTron_dns.opt.perf >
printf-4.01.0-kinFerTron_dns.opt.perf

# Summarize results
for i in *kinFerTron_dns.opt.perf; do printf "\n[$i]\n";
../../utils/crunchperf -s $i; done

On Fri, Nov 28, 2014 at 6:44 PM, Gabriel Scherer
<gabriel.scherer@gmail.com> wrote:
> Thanks for the significant effort put in documenting the bindings
> (and, of course, the cool software and research); your "information
> and documentation" page is impressive.
>
> The page has a very interesting performance comparison of numeric code
> partly or fully written in OCaml (using bigarrays of floats) -- and
> the not-so-surprising results is that the run times of the OCaml
> programs are between 100% and 200% of the run time of the reference C
> implementation.
>
> ( http://inria-parkas.github.io/sundialsml/perf.opt.png )
>
> I'm curious about this specific part of the explanation:
>
>> For instance, some OCaml versions spend a significant fraction of their time
>> in printf, and we were able to lower their ratios by instead using print_string and print_int.
>
> The new 4.02 implementation of formats, due to Benoît Vaugon, should
> be significantly faster (in my experience they match the performance of the
> less-readable print_* sequence in most situations). Did you try those
> OCaml versions with 4.02?
>
> On Fri, Nov 28, 2014 at 2:39 PM, Timothy Bourke <Timothy.Bourke@inria.fr> wrote:
>> We are pleased to announce Sundials/ML, an OCaml interface to the
>> Sundials suite of numerical solvers (CVODE, CVODES, IDA, IDAS, KINSOL).
>>
>> Information and documentation: http://inria-parkas.github.io/sundialsml/
>> Source code (BSD):             https://github.com/inria-parkas/sundialsml
>>
>> opam install sundialsml # (requires Sundials 2.5.0)
>>
>> We gratefully acknowledge the original authors of Sundials, and the
>> support of the ITEA 3 project 11004 MODRIO (Model driven physical
>> systems operation), Inria, and the Departement d'Informatique de l'ENS.
>>
>> Timothy Bourke, Jun Inoue, and Marc Pouzet.
>>



-- 
Jun Inoue