Originally posted by: SystemAdmin

Hi everybody,

I know it might sound weired but that is true;
I get the following from CPLEX.

1 of 20 MIP starts provided solutions.
MIP start 'm1' defined initial solution with objective 7559430.4380.
Tried aggregator 2 times.
MIP Presolve eliminated 2115 rows and 985 columns.
MIP Presolve modified 2180 coefficients.
Aggregator did 310 substitutions.
Reduced MIP has 2791 rows, 1567 columns, and 184928 nonzeros.
Reduced MIP has 1311 binaries, 255 generals, 0 SOSs, and 0 indicators.
Presolve time =    0.74 sec.
Clique table members: 1486.
MIP emphasis: balance optimality and feasibility.
MIP search method: dynamic search.
Parallel mode: opportunistic, using up to 8 threads.
*Segmentation fault*

any comment on that?

best
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Could you please provide additional information? Which version are you using and on which platform?
Additional, it would be great if you could send the model that produces the segfault.

One known issue is a bug in the Intel Math Kernel Library (MKL) that we are using inside the barrier algorithm (and with opportunistic parallel, we are using concurrent LP for the root node, which also calls barrier).

If you are indeed hitting the MKL bug, you could try to set the MKL_DISABLE_FAST_MM as system environment variable. This variable disables the MKL internal memory management, which fixes one of the bugs and degrades MKL performance a bit.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Hi Tobias,

Firts of all my platform is linux on a 16 core machine.
My first guess was that the number of thread might be an issue; I reduced it to 8 and it came again.
I do not have the model because I set number of threads to 1 and run it again and now is running. I am hoping for completion but if failed I will generate the model and send you.

I use SIFTING algorithm and no barrier algorithm.

Thanks for comment.

Shahin
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: RWunderling

Shahin:

Please note that sifting internally uses barrier. It does so iteratively on a subproblem consisting of a fraction of all the variables. So please do try Tobi's recommendation of using MKL_DISABLE_FAST_MM. Avoiding the problem
by only using 1 thread is consistent with this known problem in the MKL library.

Roland
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Thanks for your comments
#CPLEXOptimizers
#DecisionOptimization