topic Speed of posv and potri in Intel® oneAPI Math Kernel Library

Speed of posv and potri

xdominique — Wed, 17 Mar 2010 16:12:48 GMT

Hi!

I'm doing a bundle adjustment using posv to solve the system and potri for the covariance matrix. I use very big symmetric matrices ( > 20,000x20,000) so the calculation takes some time. The bundle adjustment is a refining process; some iterations followed by stat check accuracy forms a pass. There are mostly more than one pass for each dataset.

I recorded the time it takes for some processes (iteration, pass, solving, inversion) and I had some strange results. For small datasets (matrices smaller than 10,000), everything is ok; with the MKL library, it is very fast. :) However, with bigger ones, the time is unpredictable: be itabout computation time (MKL library) and initialisation time(program design). Doesthe computation timedepend of the exact size of the matrix, like being a multiple of 2 (which is MKL related "issue")? I read something about it in the User Manual (p. 6-14), but I wasn't sure if it applies to my problem.

Thanks for your help!

Speed of posv and potri

Chao_Y_Intel — Fri, 19 Mar 2010 06:34:43 GMT

Hello,

For large data set, does the system have enough memory to contain the data? If the data is too large and can notbe kept in the memory, swapping into disk by OS may take longer and unpredictable time.

Thanks,

Chao

Speed of posv and potri

xdominique — Fri, 19 Mar 2010 19:09:53 GMT

Hi Chao Y!

Thanks for your answer. Indeed, the matrix takes more Gb than the available RAM. So the size of the matrix being a multiple of 2 is not important?

Have a nice day.

Speed of posv and potri

TimP — Sat, 20 Mar 2010 00:39:03 GMT

I'm not familiar with the reference about multiples of 2. There is a reference about choosing sizes such that each row or column is 16-byte aligned; that should have most effect on problems much smaller than yours, while the suggestion about what fits in RAM is rightly given for larger problems.

Speed of posv and potri

Gennady_F_Intel — Tue, 23 Mar 2010 16:28:13 GMT

eskell, just for your references:

see more details about ManagingPerformanceand Memory into User's Guide, chapter 6: codingTechniques.

as an example:

To obtain the best performance with Intel MKL, ensure the following data alignment in your

source code:

Align arrays on 16-byte boundaries.

Make sure leading dimension values (n*element_size) of two-dimensional arrays aredivisible by 16.

For two-dimensional arrays, avoid leading dimension values divisible by 2048.

and etc.

--Gennady