Re: [boost] [XInt] CoW/Move Timings

Subject: Re: [boost] [XInt] CoW/Move Timings
From: Jeffrey Lee Hellrung, Jr. (jhellrung_at_[hidden])
Date: 2011-03-09 22:36:44

On 3/9/2011 5:23 PM, Chad Nelson wrote:
>
> Below are the comparative timings with and without Copy-on-Write, and
> with and without Boost.Move.
>
> Times are measured in seconds, using two sets of ten integers (512,
> 2048, and 8192 bit). The tests consisted of adding each of those pairs
> of numbers together 10,000 times, multiplying them 10,000 times, and
> doing a "mulmod" operation on them (with a third randomly-generated
> modulus argument of the same size) 1,000 times. Each test was run three
> times, under as close to identical situations as I could manage, with
> the raw results in the first three columns; the fourth column is the
> average of the three.
>
> The compiler used was GCC 4.5.1, under Ubuntu Linux 10.10. The build
> was non-debug and used -O2 optimization. Otherwise, the
> settings were left at their defaults.
>
> The "review code" lines are for the unmodified review version of the
> code. All others are for the pass-by-reference version, with several
> other modifications that should have no measurable effect on the speed.
>
> The tests show pretty conclusively that Copy-on-Write, alone or in
> conjunction with Boost.Move, provides a noticeable speed improvement
> on all three sizes. They also prove that Boost.Move can work well in
> conjunction with it. I'll leave any further analysis to people who are
> probably more qualified than I.
>
> Note that the Boost.Move numbers could be a little better. The
> integer_t classes are move-enabled, but the internal data class is not
> as yet. The data class only sometimes uses return values, so the
> difference isn't likely to be great.
>
> I will make the updated code available on request for anyone who wants
> to examine the changes.
>
> 512bit
> CoW, no Move:    0.48, 0.47, 0.47: 0.48
> No CoW, no Move: 0.61, 0.61, 0.62: 0.61 (worst)
> No CoW, Move:    0.57, 0.57, 0.55: 0.56
> CoW, Move:       0.44, 0.44, 0.44: 0.44 (best)
>
> Review code,
>    CoW, no Move:  0.50, 0.47, 0.47: 0.48
>
> 2Kbit
> CoW, no Move:    4.69, 4.79, 4.89: 4.79 (best)
> No CoW, no Move: 5.01, 5.19, 5.06: 5.09 (worst)
> No CoW, Move:    5.18, 5.00, 5.01: 5.06
> CoW, Move:       4.80, 4.79, 4.86: 4.82
>
> Review code,
>    CoW, no Move:  4.88, 4.85, 4.94: 4.89
>
> 8Kbit
> CoW, no Move:    68.0, 68.8, 68.9: 68.6
> No CoW, no Move: 70.4, 70.4, 70.2: 70.3 (worst)
> No CoW, Move:    70.9, 69.8, 69.6: 70.1
> CoW, Move:       68.4, 68.2, 67.8: 68.1 (best)
>
> Review code,
>    CoW, no Move:  68.8, 68.4, 68.8: 68.7

To echo Steven's request, please make your test code available.

The 2Kbit section has me wondering about the accuracy of these timings, 
as "CoW, no Move" outperformed "CoW, Move", which should definitely not 
be the case.  I can't think of any circumstance where move emulation 
gives you a penalty...

I think an equally important measure would be the # of allocation 
requests.  Do you track those?

- Jeff