Showing results for 
Search instead for 
Did you mean: 

Threaded MKL Lapack and performance, getrf

Hi, I've built a fortran 90 code that uses lapack's getrf and getri for inverting complex matrices. I used Intel Fortran compiler 11.1 update 5 (069) to compile the code, linking with the MKL library that came with that compiler (from what I gathered here it corresponds to the latest update to MKL. When I try to run the code with 4 threads (I have a Xeon X5570) I see the program takes 400% of the CPU, but the wall time, compared with only one thread running, is the same, or even larger! I also tried changing the size of the matrices I am inverting (from 18x18 to 100x100), but there was no change in the pattern. Am I doing something stupid here?! I can't figure it out. I searched all over Intel, and also the rest of the internet using google, and I didn't find any reference to this kind of problem. Any suggestion will be much appreciated. Cheers, Antonio
0 Kudos
1 Reply

Hi Antonio.The main problem why you don't see the advantage of running the code on 4-threads is the small input matrixes size. Intel MKL is mostly optimized for solving medium and big tasks. Could you please try to increase the size of input matrix and see the result. --Gennady