All problems are run on an Ubuntu 18.04 cluster of 3 networked machines with 64GB of RAM and Intel Xeon Gold 6130 CPUs @ 2.10GHz, yielding 32 physical cores per machine.
All jobs, including MPI jobs, are scheduled so that all 96 cores were saturated at all times.
Solver timeout is enforced by invoking octeract-engine -t, which sends a hard kill signal from the Linux shell. All time reported in the benchmarks is real time as measured by the engine using std::chrono with microsecond resolution, rounded up to the second decimal.
All problems are run using default options as described in the engine’s documentation.
– Third-party solvers
The sub-problems generated throughout the DGO procedure for this benchmark are solved using open source solvers (such as CLP and IPOPT). Because commercial solvers tend to scale better for large problems than their open source counterparts, performance for large problems can improve significantly if a commercial LP and/or NLP solver is available.