Originally posted by: PetterS

I have tried solving the following linear program:
http://www.maths.lth.se/~petter/files/fg_motion.zip (20 MB zipped MPS file)

On Windows 64-bit with Matlab interface, I had it run for ~200 iterations before I cancelled. At that point the memory usage was more than 9 GB. This memory was not reclaimed after the solver aborted gracefully.

On Linux 64 bit using the interactive solver it ran for around 40 iterations with slightly increasing memory usage before it crashed with a segmentation fault.

How can I prevent this? Is there a bug fix? I am using the academic version which I downloaded and installed today.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: PetterS

I should add that I am using the barrier method.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

This might be a known issue with the MKL library to which cplex is linked. On Linux, can you try to set environment variable MKL_DISABLE_FAST_MM to 1 and test if that cures the issue for Linux?
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Note also that the MKL issue has been fixed by Intel, and the next CPLEX release (coming this month) will include the new MKL. Consequently, your issue should be fixed by upgrading to the next CPLEX version. In the mean-time, you can either use Daniel's work-around by disabling certain features of the MKL via the environment variable, or you can just restrict the barrier algorithm to only one thread (for example, "set threads 1" in the interactive optimizer).

Tobias
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: PetterS

It seems to be working now with the environment variable. CPLEX has performed 100+ iterations without an increase in memory usage. I might add that the environment variable works well on Windows, too.

Thanks a lot for the very fast responses! I am looking forward to the next release where this is fixed.
#CPLEXOptimizers
#DecisionOptimization