Heterogeneous Cloud Computing: Design Methodology to Combine Hardware Accelerators

Heterogeneous cloud computing servers provide access to different types of hardware accelerators in order to satisfy the computational demands that standalone generalpurpose multi-processors can not deliver. The combination of different technologies provides more opportunities to accelerate the most compute-intensive applications by exploiting the key features of each type of hardware accelerator. Unfortunately, the design effort drastically increases when considering complex applications. We propose a methodology which provides an early speedup prediction when combining hardware accelerators and insights about potential performance leaks. In addition, we consider a couple of metrics to analize the computational use of the accelerators and the exploitation of the heterogeneous system. Our methodology is applied to a real-time surveillance system, which is composed of Field-Programmable Gate Arrays (FPGA) and Graphic-Processing Units (GPU). The methodology is used to provide an early speedup prediction, which is used to guide the combination of both accelerators. The achieved acceleration is only a 3:5% lower than estimated by our methodology.