A RECONFIGURABLE AND EXTENSIBLE EXPLORATION PLATFORM FOR FUTURE HETEROGENEOUS SYSTEMS

Accelerator-based -or heterogeneous- computing has become increasingly important in a variety of scenarios, ranging from High-Performance Computing (HPC) to embedded systems. While most solutions use sometimes custom-made components, most of today’s systems rely on commodity highend CPUs and/or GPU devices, which deliver adequate performance while ensuring programmability, productivity, and application portability. Unfortunately, pure general-purpose hardware is affected by inherently limited power-efficiency, that is, low GFLOPS-per-Watt, now considered as a primary metric. The many-core model and architectural customization can play here a key role, as they enable unprecedented levels of power-efficiency compared to CPUs/GPUs. However, such paradigms are still immature and deeper exploration is indispensable. This dissertation investigates customizability and proposes novel solutions for heterogeneous architectures, focusing on mechanisms related to coherence and network-on-chip (NoC). First, the work presents a non-coherent scratchpad memory with a configurable bank remapping system to reduce bank conflicts. The experimental results show the benefits of both using a customizable hardware bank remapping function and non-coherent memories for some types of algorithms. Next, we demonstrate how a distributed synchronization master better suits many-cores than standard centralized solutions. This solution, inspired by the directory-based coherence mechanism, supports concurrent synchronizations without relying on memory transactions. The results collected for different NoC sizes provided indications about the area overheads incurred by our solution and demonstrated the benefits of using a dedicated hardware synchronization support. Finally, this dissertation proposes an advanced coherence subsystem, based on the sparse directory approach, with a selective coherence maintenance system which allows coherence to be deactivated for blocks that do not require it. Experimental results show that the use of a hybrid coherent and non-coherent architectural mechanism along with an extended coherence protocol can enhance performance. The above results were all collected by means of a modular and customizable heterogeneous many-core system developed to support the exploration of power-efficient high-performance computing architectures. The system is based on a NoC and a customizable GPU-like accelerator core, as well as a reconfigurable coherence subsystem, ensuring application-specific configuration capabilities. All the explored solutions were evaluated on this real heterogeneous system, which comes along with the above methodological results as part of the contribution in this dissertation. In fact, as a key benefit, the experimental platform enables users to integrate novel hardware/software solutions on a full-system scale, whereas existing platforms do not always support a comprehensive heterogeneous architecture exploration.