Adaptive Mapping for Multiple Applications on Parallel Architectures

We propose a novel adaptive approach capable of handling dynamism of a set of applications on network-on-chip. The applications are subject to throughput or energy consumption constraints. For each application, a set of non-dominated Pareto schedules are computed at design-time in the energy, period and processors space for different cores topologies. Then, upon the starting or ending of an application, a lightweight adaptive run-time scheduler reconfigures the mapping of the live applications according to the available resources, i.e., the available cores of the network-on-chip. This run-time scheduler selects the best topology for each application and maps them to the network-on-chip using the tetris algorithm. This novel scheduling approach is adaptive, it changes the mapping of applications during their execution, and thus delivers just enough power to achieve applications constraints.