Re: [boost] [convert] Performance

Subject: Re: [boost] [convert] Performance
From: Vladimir Batov (vb.mail.247_at_[hidden])
Date: 2014-06-10 08:52:21

Matus Chochlik wrote
> On Tue, Jun 10, 2014 at 8:58 AM, Vladimir Batov <

> vb.mail.247@

> >
> wrote:
> 
>> > IMHO this could have something to do with paging or the caching of the
>> > strings.
>> > You should test the two cases by two separate programs and try to run
>> them
>> > both in the raw then cnv and cnv then raw orders.
>> > Or you could at least try to run
>> >
>> >     for (int k = 0; k < 3; ++k)
>> >         printf("str-to-int spirit: raw1/raw2=%.2f/%.2f seconds.\n",
>> >                performance_string_to_int_spirit(strings),
>> >                performance_string_to_int_spirit(strings));
>> >
>> > to try it out.
>>
>> Matus,
>>
>> Geez, I would never figure that out by myself. Thanks. So, the first call
>> would pull the strings into the fast cache so that the second call would
>> not
>> have to, right? Now I changed the test as
>>
>>     int const num_tries = 20;
>>     double     raw_time = 0;
>>     double     cnv_time = 0;
>>
> 
> Vladimir,
> 
> if this is indeed caused by caching and you want to run both tests in a
> single program, then you should do a "dry run" that will cause the strings
> to be cached before you do the actual timing.
> Also I'd increase the number of tries at least to 200.
> 
>>
>>     for (int k = 0; k < num_tries; ++k)
>>         raw_time += performance_string_to_int_spirit(strings);
>>
>>     for (int k = 0; k < num_tries; ++k)
>>         cnv_time += performance::str_to_int(strings,
>> boost::cnv::spirit());
>>
>>     printf("str-to-int spirit: raw/cnv=%.2f/%.2f seconds.\n",
>>            raw_time / num_tries,
>>            cnv_time / num_tries);
>>
>> Does it look fair to you? It surely gives more realistic results with
>> bosst::convert() adding 0.7% overhead:
>>
>> 32:~/dev/boost.convert.>> ~/dev/bin/test-convert
>> str-to-int spirit: raw/cnv=1.41/1.42 seconds.

Well, it seems I spoke too soon. I was rushing out and tried the above in a
hurry and, as a result, messed it up. Now I am doubtful it's cache-related
as I simply run

    for (int k = 0; k < num_tries; ++k)
    {
        double raw_time1 = performance_string_to_int_spirit(strings);
        double raw_time2 = performance_string_to_int_spirit(strings);

        printf("str-to-int spirit: raw1/raw2=%.2f/%.2f seconds.\n",
               raw_time1, raw_time2);
    }

and both raw_time1 and raw_time2 are the same (well, there is some variation
but really minor and random). Then I run 

    for (int k = 0; k < num_tries; ++k)
    {
        double cnv_time = performance::str_to_int(strings,
boost::cnv::spirit());
        double raw_time = performance_string_to_int_spirit(strings);

        printf("str-to-int spirit: raw/cnv=%.2f/%.2f seconds (%.2f%%).\n",
               raw_time, cnv_time, 100 * cnv_time / raw_time);
    }

and regardless if I calculate cnv_time first or raw_time first I get the
same results (i.e. the first call does not help the second call at all) as
below: 

str-to-int spirit: raw/cnv=1.46/1.42 seconds (97.34%).
str-to-int spirit: raw/cnv=1.47/1.42 seconds (96.15%).
str-to-int spirit: raw/cnv=1.46/1.42 seconds (97.31%).
str-to-int spirit: raw/cnv=1.45/1.42 seconds (97.45%).
str-to-int spirit: raw/cnv=1.45/1.41 seconds (97.30%).

I give up. My brain is too puny to understand that. 

--
View this message in context: http://boost.2283326.n4.nabble.com/review-Convert-library-tp4662821p4663936.html
Sent from the Boost - Dev mailing list archive at Nabble.com.