On the impact of dynamic data management for distributed local memories in heterogeneous MPSoCs

Abstract

With the increasing number of integrated functional units in Multi-Processor System-on-Chip (MPSoC) the communication among modules is becoming a major challenge. The overall system performance is not only characterized by the computing power but more often limited by the slow interconnections and memory accesses. Furthermore, the available on-chip memory is not efficiently utilized according to the application requirements. This paper describes the concept, implementation and analysis of a data management unit, which improves the utilization of on-chip distributed local memories. Hence data locality and system performance is improved. This strategy leads to a dramatic reduction in the number of accesses to the external memory for data-intensive applications. The impact of this approach on system performance is investigated in this work, which shows a reduction of 53% in the number of accesses to the external memory. In data-limited environments this leads to an improvement of up to 30% in the overall system performance.