Fast and Energy-Driven Design Space Exploration for Heterogeneous Architectures

Abstract

In the last years, the integration of specialized hardware accelerators in Multiprocessor System-on-Chip (MpSoC) led to a new kind of architectures combining both software (SW) and hardware (HW) computational resources. For these new Heterogeneous MpSoC (HMpSoC) architectures, performance and energy consumption depend on a large set of parameters such as the HW/SW partitioning, the type of HW implementation or the communication cost. Design Space Exploration (DSE) consists in adjusting these parameters while monitoring a set of metrics (execution time, power, energy efficiency) to find the best mapping of the application on the targeted architecture. With the shift from performance-aware to energy-aware designs, computer-aided design and development tools try to reduce the large design space by simplifying HW/SW mapping mechanisms. However, energy consumption is not well supported in most of DSE tools due to the difficulty to fast and accurately estimate the energy consumption. To this aim, this work introduces a DSE method based on an analytical power model to circumvent the computation time bottleneck of state-of-the-art DSE methods. This exploration method proposes to optimize the HW/SW partitioning and mapping under user-defined objectives, especially an energy constraint. It targets tiling-based parallel applications and relies on an analytical power model that provides the DSE framework with the execution time and energy of a HW/SW configuration. The power model parameters are obtained with the measurements of a tiny subset of the design space, which are then injected into two extraction functions to obtain analytical formulations of the execution time and the energy consumption of the computation kernel. The partitioning problem constraints are defined as a set of inequalities with Boolean, integer (discrete) and non-integer (continuous) variables within a Mixed Integer Linear Programming (MILP) framework. Then, the best configuration that minimizes the user objective (e.g. execution time or total energy consumption) can be efficiently determined using commercial or open source solvers within a second. This methodology was tested on a Zynq-based heterogeneous architecture with two application kernels: a matrix multiplication and a Stencil computation. The results show a minimum of 12% acceleration speed-up and energy saving compared to standard approaches. They also show that the most energy-efficient solution is application-and platform-dependent and moreover hardly predictable. Such method could be included in a complete framework with a multi-step exploration to obtain an energy-efficient mapping of a full application on HMpSoC and to open new opportunity for future computer-aided design tools.