Difference in results while performing matrix multiplication

[nvthuyne1]

Hello all,

I reached out to the R-devel community for a question, but it didn't get accepted. So I will try again here. It is concerning differences in matrix multiplication for large matrices using different blas libraries on different environments.
As I am representing a rather large user community at BASF and this is critical for our day to day work, I would really appreciate any light that could be shed on this topic.

With kind regards,

r-devel-owner@r-project.org|9:40 AM (18 minutes ago)
to me

The message's content type was not explicitly allowed

---------- Forwarded message ----------
From: ... <[...)>
To: r-devel@r-project.org
Cc:
Bcc:
Date: Fri, 29 Mar 2019 15:15:51 +0100
Subject: Difference in results while performing matrix multiplication

Hello R-devel,

While performing some migration from cluster A to B, one of my users has noticed something strange in the matrix multiplication.

The old environment:

sessionInfo()
R version 3.4.1 (2017-06-30)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Red Hat Enterprise Linux Server release 6.9 (Santiago)

Matrix products: default
BLAS: /gpfs/gssgpfs1/biogrid/tools/bioinfo/app/R-3.4.1/lib64/R/lib/libRblas.so
LAPACK: /gpfs/gssgpfs1/biogrid/tools/bioinfo/app/R-3.4.1/lib64/R/lib/libRlapack.so

the new environment

sessionInfo()
R version 3.5.1 (2018-07-02)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Red Hat Enterprise Linux Server 7.4 (Maipo)

Matrix products: default
BLAS/LAPACK: /gpfs/gssgpfs1/biogrid/tools/eb/software/OpenBLAS/0.3.1-GCC-7.3.0-2.30/lib/[libopenblas_haswellp-r0.3.1.so](http://libopenblas_haswellp-r0.3.1.so/)

This is clearly the difference: BLAS/LAPACK

When performed on the old environment we get the following:

On the new environment we get this

To make matters even worse, the result changes at each run, so this can't be correct right?

Could anybody shed some light here on what might be going on?

with kind regards,

[EconomiCurtis]

I don't know anything about BLAS/LAPACK, so definitely take my reply with a grain of salt. But perhaps these new methods rely on some kind of sampling methodology? That could explain the differences between these methods and with the new method, between multiple identical matric multiplications.

You might check out with a forum that specializes in these methods,
http://icl.cs.utk.edu/lapack-forum/viewforum.php?f=3

[nvthuyne1]

Thanks EconomiCurtis, I will have a look at the forum.

In the meantime, I have added the data from my example to a github repository

GitHub - nvthuyne/RData: data for issue with Lapack/Blas Library

This might give other people the possibility to try this issue for themselves.
I really hope somebody can help me here, as this might be a breaking issue for the user community