For the widespread application of nonlinear model-predictive control (NMPC) in the chemical industry, the computational effort that is required for the solution of the underlying resulting nonlinear dynamic optimization problems is a major obstacle. For complex process models and long prediction and control horizons, the computation times lead to large sampling times and feedback delays, which may cause performance deterioration and constraint violations. A promising approach to speed up the computations has been proposed in . In this approach, two main ideas are put forward: On the one hand, the structure of the quadratic programming (QP) subproblems that result from using Sequential Quadratic Programming for the solution of the NLP problems is exploited, and different approximations of the exact solution are used in a parallel and hierarchical fashion (multi-level iterations). On the other hand, in each iteration of the algorithm the QP problems can be divided into a preparation phase based on the previous solution and a feedback phase based on the current state of the process (real-time iterations). In the present work we employ parallel computation to investigate the effects of computational delays and realize the parallel execution of two different approximations of the multi-level modes to reduce the computational burden. We use a benchmark semi-batch polymerization reactor model to investigate and to compare the control performances for an NMPC scheme with an economically motivated cost function using a standard implementation of an online optimizing control scheme, a scheme using real-time iterations and a scheme using multi-level iterations in a parallelized and interleaved fashion.