forum.alglib.net

Hello,

I'm new with alglib and i want to know if its possible to use the method rmatrixinverse with multithreading to perform the operations inversing the matrices. I was looking for documentation but i didnt find nothing about it.

Thanks for your responses

1. in the current ALGLIB version - no, it is not possible. In the upcoming 3.9.0 you will be able to do so.

2. However, nonsymmetric matrix inverse is much harder to parallelize than, say, matrix-matrix product. In our experiments significant speedups from multithreading were achieved only for N's larger than 1000.

Thanks a lot for your kickly response, i'm working with symetric matrices between 500 and 1500 N's as part of a linear system of type Ax=b. Can i do something with the current version for optimice the calcullation of the inverse?.

Thankyou for the help

If we assume that performance of C# version is "one unit" then for inversion of positive-definite matrices with N=1024 we have following performance metric:
* ALGLIB for C# performance = 1.0
* ALGLIB for C++ (generic C) performance = 3.1x
* commercial ALGLIB single-threaded performance = 5.4x
* commercial ALGLIB 5-threaded performance = 13.7x = 2.5 * performance of single-threaded version

You may see that switching to high-performance computational core give significant speed-up. Multithreading gives significant speed-up for N=1024, but still sublinear. Of course, these metrics are related only to SPD matrix inversion, and are valid only for 3.8.2 and maybe 3.9.0.

P.S If you use commercial edition of ALGLIB, I may send you a quick patch which will enable multithreaded acceleration for SPD matrices.

Hi again,

I was cheking my matrices but they are not SPD, so i supose that i must try another thing or wait for the 3.9 version. Do you have any idea of the date of publication of the new version?. Thank you very much for your help.

3.9.0 is in the final stages of development. Several weeks, I hope...

Thanks i'll wait for it ;)