Implementation of Concurrent Parallelization of Branch-and-bound algorithm in Everest Distributed Environment

Effectively parallelizing of branch-and-bound (BnB) solvers for mathematical programming problems with discrete variables in distributed computing environment is an important issue. Recently Everest web-based environment has been equipped with a generic message service. This improvement enables BnB-solvers to run in parallel and to exchange incumbents values. Thus Everest environment got generic application that parallelizes BnB-solvers (SCIP and COIN-OR CBC) in a rather simple way. Initially, the application was based on preliminary decomposition of feasible domain, e.g. by fixing values of some integer variables. In the current article further improvement of this Everest application is presented. It supports so called “concurrent parallelization” when all BnB-solvers work with the same optimization problem but with different settings defining search path in BnB-tree. Thus, running solvers have more chance to find incumbents, to exchange their values and to reduce BnB search. Preliminary testing of improved BnB-service on Traveling Salesman Problem shows noticeable speedup.