Taming extreme heterogeneity via machine learning based design of autonomous manycore systems
To avoid rewriting software code for new computer architectures and to take advantage of the extreme heterogeneous processing, communication and storage technologies, there is an urgent need for determining the right amount and type of specialization while making a heterogeneous system as programmable and flexible as possible. To enable both programmability and flexibility in the heterogeneous computing era, we propose a novel complex network inspired model of computation and efficient optimization algorithms for determining the optimal degree of parallelization from old software code. This mathematical framework allows us to determine the required number and type of processing elements, the amount and type of deep memory hierarchy, and the degree of reconfiguration for the communication infrastructure, thus opening new avenues to performance and energy efficiency. Our framework enables heterogeneous manycore systems to autonomously adapt from traditional switching techniques to network coding strategies in order to sustain on-chip communication in the order of terabytes. While this new programming model enables the design of self-programmable autonomous heterogeneous manycore systems, a number of open challenges will be discussed.
