We develop a novel framework for optimized composition of explicitly parallel software components with different implementation variants given the problem size, data distribution scheme and processor group size on a Linux cluster. We consider two approaches (or two cases of the framework).  In the first approach, dispatch tables are built using measurement data obtained offline by executions for some (sample) points in the ranges of the context properties. Inter-/extrapolation is then used to do actual variant-selection for a given execution context at run-time. In the second approach, a cost function of each component variant is provided by the component writer for variant-selection. These cost functions can internally lookup measurements' tables built, either offline or at deployment time, for computation- and communication-specific primitives. In both approaches, the call to an explicitly parallel software component (with different implementation variants) is made via a dispatcher instead of calling a variant directly. As a case study, we apply both approaches on a parallel component for matrix multiplication with multiple implementation variants. We implemented our variants using Message Passing Interface (MPI). The results show the reduction in execution time for the optimally composed applications compared to applications with hard-coded composition. In addition, the results show the comparison of estimated and measured times for each variant using different data distributions, processor group and problem sizes.