topic Nondeterminism in Pardiso in Intel® oneAPI Math Kernel Library

Nondeterminism in Pardiso

Noel_B_ — Tue, 12 Oct 2010 20:13:41 GMT

Hi,

We use the Intel MKL Pardiso solver to solve a small (couple hundred row) sparse (4%) system. Running our test case two consecutive times using the Pardiso solver produces very different results. In fact, there seems to be inherent non-determinism in the Pardiso solver.

Now for the standard caveats, our code is memory clean, both purify and valgrind are content. Replacing Pardiso with a standard biconjugate gradient solver and running the same test case on the same hardware produces identically repeatable results. We see this hysteresis on Windows, Mac and Linux in both debug and -O2. The code is serial, we don't enable OpenMP nor do we use the parallel Pardiso solver.

Has anyone else noticed this behavior or is this a known potential problem, any ideas? We haven't yet tried to produce a small test case demonstrating this problem, thought we'd ask first.

-- Noel Belcourt

S896822:~ kbelco$ icpc -V

Intel C++ Intel 64 Compiler Professional for applications running on Intel 64, Version 11.1 Build 20100806 Package ID: m_cproc_p_11.1.089

Nondeterminism in Pardiso

Gennady_F_Intel — Wed, 13 Oct 2010 04:52:29 GMT

Might beyour matrices areill-conditioned or close to them. Did you estimate the conditin numbers of your input matrixes?

--Gennay

Nondeterminism in Pardiso

barragan_villanueva_ — Wed, 13 Oct 2010 06:33:05 GMT

Please look at similar thread

Nondeterminism in Pardiso

basel — Wed, 13 Oct 2010 11:36:43 GMT

Hi Noel,

An answer to your question might be that even in case of iparam(2)=1 (#cores) you must use single threaded BLAS.

As an alternative you could also use the new version PARDISO 4.0.0 (www.pardiso-project.org). This version will always compute a bit-by-bit identical solution even on multicores e.g. using eight cores for the parallel factorization and the parallel solution.

Regards,
Olaf Schenk

Nondeterminism in Pardiso

Noel_B_ — Wed, 13 Oct 2010 20:19:47 GMT

I love answering my own question.

Our matrix is not ill-conditioned. We did find the source of the non-determinism. We were passing a phase=0 in our last call to pardiso to release internal memory. Phase=0 releases memory for L and U. Passing phase=-1 fixed our problem. Phase=-1 releases all internal memory for all matrices. Changing phase from zero to negative one eliminated the nondeterminism in the solver.

I'd like to request that the developers look at this issue as we nearly gave up on Pardiso entirely due to the inability to reproduce the exact same answer given the exact same inputs. Perhaps some additional (re) initialization can be added with phase=0 to ensure both phases (0 and -1) produce the exact same answer.

Thanks for your responses.

-- Noel Belcourt

Nondeterminism in Pardiso

Noel_B_ — Wed, 13 Oct 2010 21:18:27 GMT

As an alternative you could also use the new version PARDISO 4.0.0 (www.pardiso-project.org). This version will always compute a bit-by-bit identical solution even on multicores e.g. using eight cores for the parallel factorization and the parallel solution.

In fact, we will try this out as eliminating sources of non-determinism in parallel are extremely important to us, e.g. MPI and OpenMP reductions in rank order and so forth.

Look forward to trying your solver.

-- Noel Belcourt