Performance Issues

[ahassan]

I compiled CLAPACK 3.0 using Microsoft visual studio 6.0 C++ compiler.
The OS is windows XP Professional.
The machine is Pentium 4 with 512 MB RAM.

The performance is not good at all. It took 4 seconds to do LU factorization on 1000*1000 matrix, while matlab took .5 second on the same machine.

If anyone can help in this issue, he/she can email me on alihuco@yahoo.com. I highly appreciate that.

One solution might be using Intel Math Kernel Library. Please direct me to an easy tutorial on how to use MKL from MVS VC++ if possible. An example project will be highly appreciated.

Thanks.
Ali

[Julie]

Dear Ali,

The performance of LAPACK depends mainly on the BLAS library you use.
I guess that you are using a non-optimized BLAS with CLAPACK and using the optimized blas from MATLAB. This explains why CLAPACK is so slow.

You can try to link CLAPACK with the MKL library, it is very easy (In fact you just need to locate where your mkl library is on your computer and add the lib to the linking sequence)
Here is the help from Intel: http://www.intel.com/support/performanc ... 017282.htm

Let me know if you need some more help.
Sincerely
Julie Langou

[ahassan]

Dear Mr. Langou,

I appreciate your prompt response. I used the info. You provided with trial version of MKL 9.0 from Intel. It is amazing (very fast).

I am a researcher in the middle of a math-computer science project. So please know that during the progress of my project, I will probably need to ask more questions.

For now, is there any way to get MKL performance for free?


Thank you very much.
ali

[Julie]

Dear Ali,

Glad you got your code running fine and faster.
Yes, there are some "free" optimized blas implementation that will give you very good performance.

The two principal are :
- atlas http://math-atlas.sourceforge.net/
- goto blas http://www.tacc.utexas.edu/resources/software/#blas

I don't really know which one is easier to install on Windows platform.
Good luck
Julie