Nondeterminism
Among the ILOG CPLEX parallel optimizers, only parallel simplex follows a deterministic algorithm, producing the same number of iterations and the same solution path when you apply it to the same problem more than once. In contrast, the parallel barrier and parallel MIP optimizers are nondeterministic: repeated solutions of a model using exactly the same settings can produce different solution paths and, in the case of the parallel MIP optimizer, very different solution times and results.
The basic algorithm in the ILOG CPLEX Parallel MIP Optimizer is branch & cut. The primary source of parallelism in branch & cut is the solution of the LP or QP subproblems at the individual nodes of the search tree. These subproblems can be distributed over available processors to be carried out in parallel. The individual solution paths for these subproblems will, in fact, be deterministic, but the speed at which their solutions occur can vary slightly. These variations lead to nodes being taken from and replaced in the branch & cut tree in different order, and this reordering leads to nondeterminism about many other quantities that control the optimization. This nondeterminism is unavoidable in such a context, and its effects can result in some cases in very different solution paths.
