Re: LuaJIT2 performance for number crunching

[Francesco Abbate <francesco.bbt@...>]

2011/2/16 Leo Razoumov <slonik.az@gmail.com>:
> Wow! This is impressive. BTW, is it possible to make the test case run
> longer, let say, 5 seconds for Lua.
> CPU throttling causes large timing variability of very short processes.

Hi Leo,

I've already made this test, I've changed t1 to 2000 instead of 200
and the benchmark was confirmed, LuaJIT2 is faster then C code using
the GSL library and compiled with -O2.

I've made again the tests right now in my windows computer at work.
Here the results:

rk4-unroll.lua with LuaJIT2: 0m4.391
C code with GSL lib (opt -O2): 0m20.094

I don't know why in my windows system the factor is even greater 4.5x
improvement instead of 2.6x I've obtained on my linux box at home. I
consider linux results more representative since I don't have stupid
corporate antivirus running all the time in the background :-)

Otherwise, I've made some progress in the ODE integration
implementation. Thanks to Steve Donovan I've now a nice template
engine for the code to generalize the integration routine to a system
of dimension N.

I'm working now to implement the system to control the error and
adjust the step of automatically. This is something required in real
cases where a fixed step size would be a too much coarse approach.

-- 
Francesco