variability in results from 32 vs 64 bit PARDISO

[tyr]

Hi,

Should I expect to get identical results ( up to the 15th decimal place ) from a 32-bit and 64-bit compilation of my calculation when using single-threaded PARDISO via DSS interface with identical inputs ( the matrix and rhs are identical up to 15 decimal places )?

I am using sequential MKL (v 10.2.5) so there is no threading. My matrix is small, 108x108; its cond # is on the order of 10^3. I am observing rounding/truncation differences in the lhs between the 32-bit and 64-bit builds of otherwise the same source. Using mkl_alloc() does not affect the differences I observe. The only thing that is different between the builds that i'm aware of are the mkl libraries for the two platforms. Is it possible that there can be rounding/trunctation differences between 32-bit and 64-bit PARDISO solver?

Thx, tyr

[Chao_Y_Intel]

Hi tyr,

Some variation in output is fine. The following article discussed the reason on this:

http://software.intel.com/en-us/articles/getting-reproducible-results-with-intel-mkl/

If such difference are the incorrect results, we need to have a test code to investigation.

Thanks,
Chao

[TimP]

Quoting Chao Y (Intel)

Some variation in output is fine. The following article discussed the reason on this:

http://software.intel.com/en-us/articles/getting-reproducible-results-with-intel-mkl/

In particular, that article mentions the question of alignments, which probably vary between 32- and 64-bit OS, if you don't apply suggested methods to control them.

[TimP]

Quoting Chao Y (Intel)

Some variation in output is fine. The following article discussed the reason on this:

http://software.intel.com/en-us/articles/getting-reproducible-results-with-intel-mkl/

In particular, the alignments mentioned there are almost certain to vary between 32- and 64-bit OS (more so on Windows than linux), if you don't specify how you want them handled.