Keywords: rendering, blender cycles, energy efficient computing, MERIC, high performance computing.

Image rendering is an example of a task, that usually runs for a long time and it can easily utilize a whole supercomputing cluster, therefore it consumes a lot of energy. Even minimal power reduction can result in significant energy and money savings, as well as reduced carbon footprint. So far, the main focus concerning the energy consumption in image rendering has concentrated mainly on mobile platforms. Energy consumption of computationally extensive tasks is crucial on mobile devices because it directly influences battery life. There are several related researches in energy efficient High Performance Computing (HPC) working on reduction of system resources while conserving the application’s time to solution. Our observation is that the consumed energy might be reduced despite extending the application runtime. For such case we cannot consider just the energy consumed by the CPUs but instead evaluate the consumption of the whole computational node. In our contribution we evaluate the energy consumption optimization of image rendering on a typical architecture of an HPC system. As a renderer we use CyclesPhi, which is our own modified version of Cycles renderer from Blender 3D creation suite. CyclesPhi fits the HPC environment in such a way, that it runs as a client on one or multiple nodes and efficiently utilizes the cluster by optimal load balancing. In order to reduce energy consumption of a scene rendering we have used MERIC, our own developed library for HPC application profiling and runtime tuning. MERIC is searching for configuration of hardware, system software, and application parameters which can provide minimal energy consumption for each manually instrumented region inside the analysed application. For the production runs of the application, MERIC then sets the optimal configuration during the application runtime. In this way we have instrumented the Blender client and analysed the rendering task. On Haswell architecture (two Intel Xeon E5-2680v3 processors per node) we were able to save 9% of energy while extending the rendering time by 21%. If more conservative setting was applied, we have saved 4.8% of energy and prolonged the rendering time by 4%.

download the full-text of this paper (PDF, 2604 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £25 +P&P)