Efficient Reconfigurable Global Network-on-Chip Designs towards Heterogeneous CPU-GPU Systems: An Application-Aware Approach

Abstract

Different applications require different communication performance between subnets in a global hybrid network-on-chip (NOC) of a heterogeneous CPU-GPU architecture (HSA). It is impractical to deploy (at design time) or switch-on (at runtime) all the hybrid routers in the network for a certain application that needs several hybrid routers for communication. Reconfiguring the customized global hybrid NOC is important because the cost of deploying/powering these hybrid routers will be reduced significantly when the network scales up. Hence, applying optimization is feasible on this problem.We consider the problem of optimizing the quantity of the utilized hybrid routers in the global hybrid NOC, when the applications and network configurations are known. This problem can be cast as a mixed-integer linear programming which is known to be NP-Hard in general.We propose a prediction model of estimating the near-optimal amount of the utilized hybrid routers with a quick time. Our evaluation shows that the solution time of the prediction model outperforms that of the conventional model by 99 percent on average. The models also saved up to 84 percent on average in terms of the router utilization, compared to without using the models. We validated our estimated results by simulating them in HSA to prove the efficient performance.