The paper concerns task scheduling in dynamic SMP clusters based on the notion of moldable computational tasks. Such tasks have been used as atomic elements in program scheduling algorithms with warranty of schedule length. For program execution, a special shared memory system architecture is used. It is based on dynamic processor clusters, organized around shared memory modules by switching of processors between memory module busses. Fast shared data transfers between processors inside such clusters can be performed through data reads on the fly. The dynamic SMP clusters are implemented inside system on chip (SoC) modules additionally connected by a central global network. A task scheduling algorithm is presented for program macro dataflow graphs for execution in the assumed architecture. The algorithm first identifies a set moldable tasks in a given program graph. Next, this set is scheduled using a 2-phase algorithm including allotment of resources to moldable tasks and final list scheduling, with a warranty of schedule length. The complete algorithm has been implemented as a program package and examined using simulated execution of scheduled program graphs.