A neuro-evolution method for dynamic resource allocation on a chip multiprocessor

Technology-driven limitations will soon force microprocessor chips to contain multiple processing cores, as the scalability of individual cores peaks while transistor counts continue to increase. To obtain the best performance, flexible management of the on-chip resources, such as cache memory and off-chip bandwidth, is needed. However, the control for the dynamic management of these on-chip resources is difficult to design. In this paper, we propose a method for developing such a controller: evolving a recurrent neural network using the enforced sub-populations algorithm. The method is tested in a trace-based simulation that measures dynamic assignation of a pool of level-two cache banks to a set of processing cores. We present results showing that, when the chip is controlled by the neural network, we obtain a 13% performance improvement over static cache partitioning.