Molecular simulation has become a standard tool for studying the function of biomolecules, such as proteins, nucleic acids and lipids. Due to increasing computer power and decreasing length scales in engineering, molecular simulation is also increasingly used in microfluidics and the study of, for instance, small water droplets. All these applications would benefit strongly from simulations that are several orders of magnitude longer than the current state of art. Although currently Moore's law still holds, the performance of processor cores no longer doubles every 18 months, but rather the number of cores increases. Therefore to improve the performance and to scale to a million cores, each core should do less work. With the classical single-program multiple-data parallelism the communication time will quickly become a bottleneck. To advance the molecular simulation field and efficiently use upcoming million core computers, a switch to multiple-program multiple-data parallelism (MPMD) is required. Domain decomposition should be applied over the nodes, whereas within a node MPMD parallelism should be used. This requires workloads being divided and dispatched efficiently to different threads. To hide the communication times, calculation should be overlapped with communication. Because simulation time steps will soon take in the order of 100 microseconds, global communication will become a bottleneck. However,global communication is required for, among other things, full electrostatics algorithms. Thus new algorithms need to be derived to ensure parallel scaling. Only with such efforts we will be able to fully utilize the potential of upcoming hardware to solve current and future scientific problems.