Segmentation fault solving large linear program

[Yogesh Dahiya]

Hi,

I am trying to solve a large linear program where A (in Ax<=b) is of order of 500000*100 and I am getting segmentation fault with default KKT sovler. Backtrace using gdb points it to dscal_k () in libopenblas.so . My cvxopt version is 1.1.9

 and I have installed it via conda. Any thoughts how I can get it working. Thanks in advacne.

[Martin]

Did you install CVXOPT via the conda-forge channel? 

There was recently a post with a similar issue with OpenBLAS on a system with many cores. I believe that in that case the problem was due to a bug in OpenBLAS—setting the environment variable OPENBLAS_NUM_THREADS=1 supposedly worked. Perhaps you could try that? Alternatively, you may want to build CVXOPT with a different BLAS/LAPACK library.

[Yogesh Dahiya]

Yes I installed using conda-forge channel and OPENBLAS_NUM_THREADS=1 works for me as well. Does this restrict performance in any way? Thanks

[Martin]

Yes, it may have an impact on performance, especially if you solve large problems and have a CPU with many cores. It appears that the next release of OpenBLAS is just a few weeks away (see https://github.com/xianyi/OpenBLAS/issues/1223), and hopefully it will fix the bug that gives rise to this issue.

An alternative would be to build CVXOPT with MKL or ATLAS instead of OpenBLAS.