Skip to Content
0
Former Member
May 08, 2012 at 12:36 AM

Parallel CPU usage with SNP Optimizer

123 Views

Hello, I would like to know if it is now possible to execute SNP optimization runs that will utilize more than one CPU on the Opt Server. Previously an optimization run was limited to being solved only on one CPU, but I have read that recent ILOG Cplex releases support the solving of an optimization run on multiple CPUs thereby significantly reducing runtimes. I have not been able to find any information online from SAP on whether this capability has now been enabled with the SNP optimizer. This capability would much improve solve times for discrete problems.

Question

For what class of problems can parallel CPLEX improve performance?

Answer

Parallel CPLEX can significantly improve performance for certain classes of problems. It works particularly well with:

  • MIPs, especially with CPLEX 12.2 or later;
  • barrier method for LPs, QPs and QCPs;
  • concurrent optimization for LPs and QPs.

With MIPs, CPLEX processes nodes in the branch and bound tree on separate threads. Adding multiple processors or cores can cut the run time in half or better on some models. CPLEX 12.2 and later also perform parallel processing at the root node. However, for earlier versions of CPLEX, the majority of the MIP runtime needs to be spent processing nodes for Parallel CPLEX to improve performance.

For LPs and QPs, the barrier optimizer of CPLEX runs in parallel. The barrier optimizer runs well in parallel because most of the computations are focused on the solution of a single linear system of equations.

For LPs and QPs, CPLEX also offers a concurrent optimization procedure, where CPLEX solves your model with a different LP algorithm on a different processor, stopping as soon as the first one has solved the problem. This feature is not available for QCPs since CPLEX currently only uses the barrier method to solve such models. Note that CPLEX 12.2 and later use concurrent optimization for the root node solve by default, thus potentially improving the run time for MIPs where the root node solve comprises a significant percentage of the total solve time.