Manycore architectures are expected to dominate future general-purpose and application-specific computing systems. The ever-increasing number of on-chip processor cores and the associated interconnect complexities present significant challenges in the design, optimization and operation of these systems. In this paper we investigate the applicability of intelligent, dynamic system-level optimization techniques in addressing some manycore design challenges such as dynamic resource allocation. In particular, we introduce hardware enabled system-level bidding-based algorithms as an efficient and real-time on-chip mechanism for resource allocation in homogeneous and heterogeneous (MPSoC) manycore architectures. We have also developed a low-level simulation framework, to evaluate the proposed bidding-based algorithms in several on-chip network-connected manycore configurations. Experimental results indicate performance improvements between 8%–44%, when compared to a standard on-chip static allocation, while achieving a balanced workload distribution. The proposed hardware was synthesized to show that it imposes a very small hardware overhead to the overall system. Power consumption of the embedded mechanism as well as energy consumption due to additional network traffic for collecting system feedback are also estimated to be very small. The obtained results encourage further investigation of the applicability of such intelligent, dynamic system-level algorithms for addressing additional issues in manycore architectures.
To access the full article, please see PDF.