Re: [boost] [units] runtime performance again

From: Matthias Schabel (boost_at_[hidden])
Date: 2007-04-04 01:46:13

> I added numeric solution of a simple differential equation to the  
> benchmark.
>
> output from 3 runs:
>
> multiplying ublas::matrix<double>(1000, 1000) : 16.016 seconds
> multiplying ublas::matrix<quantity>(1000, 1000) : 16.453 seconds
> tiled_matrix_multiply<double>(1000, 1000) : 1.859 seconds
> tiled_matrix_multiply<quantity>(1000, 1000) : 1.843 seconds
> solving y' = 1 - x + 4 * y with double: 3.219 seconds
> solving y' = 1 - x + 4 * y with quantity: 2.656 seconds
>
> multiplying ublas::matrix<double>(1000, 1000) : 16.281 seconds
> multiplying ublas::matrix<quantity>(1000, 1000) : 16.406 seconds
> tiled_matrix_multiply<double>(1000, 1000) : 1.906 seconds
> tiled_matrix_multiply<quantity>(1000, 1000) : 1.859 seconds
> solving y' = 1 - x + 4 * y with double: 3.281 seconds
> solving y' = 1 - x + 4 * y with quantity: 2.61 seconds
>
> multiplying ublas::matrix<double>(1000, 1000) : 16.094 seconds
> multiplying ublas::matrix<quantity>(1000, 1000) : 16.516 seconds
> tiled_matrix_multiply<double>(1000, 1000) : 1.875 seconds
> tiled_matrix_multiply<quantity>(1000, 1000) : 1.859 seconds
> solving y' = 1 - x + 4 * y with double: 3.203 seconds
> solving y' = 1 - x + 4 * y with quantity: 2.672 seconds

Interesting - this definitely shows some of the pitfalls of simple  
performance testing. Here are my results :

multiplying ublas::matrix<double>(1000, 1000) : 42.78 seconds
multiplying ublas::matrix<quantity>(1000, 1000) : 42.31 seconds
tiled_matrix_multiply<double>(1000, 1000) : 1.73 seconds
tiled_matrix_multiply<quantity>(1000, 1000) : 2.05 seconds
solving y' = 1 - x + 4 * y with double: 2.59 seconds
solving y' = 1 - x + 4 * y with quantity: 2.57 seconds

multiplying ublas::matrix<double>(1000, 1000) : 42.74 seconds
multiplying ublas::matrix<quantity>(1000, 1000) : 42.43 seconds
tiled_matrix_multiply<double>(1000, 1000) : 1.77 seconds
tiled_matrix_multiply<quantity>(1000, 1000) : 2.08 seconds
solving y' = 1 - x + 4 * y with double: 2.59 seconds
solving y' = 1 - x + 4 * y with quantity: 2.59 seconds

multiplying ublas::matrix<double>(1000, 1000) : 42.78 seconds
multiplying ublas::matrix<quantity>(1000, 1000) : 42.31 seconds
tiled_matrix_multiply<double>(1000, 1000) : 1.74 seconds
tiled_matrix_multiply<quantity>(1000, 1000) : 2.06 seconds
solving y' = 1 - x + 4 * y with double: 2.6 seconds
solving y' = 1 - x + 4 * y with quantity: 2.57 seconds

I'm not sure why my simple matrix multiplication results are so much  
slower...the others are comparable. In any case, the relative  
performance is obviously close enough to identical for non-HPC  
applications... I think the performance arguments look like red  
herrings... I'll put this in the sandbox example code.

Matthias