Multi-Objective Local Instruction Scheduling for GPGPU Applications

Constantin Timm, Frank Weichert, Peter Marwedel, and Heinrich Müller


Power-Aware Systems, Compilers and Run Time Support, GPGPU, Multi-Objective Genetic Algorithm


In this paper, a new optimization approach (MOLIS: Multi-Objective Local Instruction Scheduling) is presented which maximizes the performance and minimizes the energy consumption of GPGPU applications. The design process of writing efficient GPGPU applications is time-consuming. This disadvantage mainly arises from the fact that the optimization of an application is accomplished in an expensive trial-and-error manner without efficient compiler support. Especially, efficient register utilization and load balancing of the concurrently working instruction and memory pipelines were not considered in the compile process up to now. Another drawback of the state-of-the-art GPGPU application design process is that energy consumption is not taken into account, which is important in the face of green computing. In order to optimize performance and energy consumption simultaneously, a multi-objective genetic algorithm was utilized. The optimization of GPGPU applications in MOLIS employs local instruction scheduling methods. The optimization potential of MOLIS was evaluated by profiling the runtime and the energy consumption on a real platform. The optimization approach was tested with several real-world benchmarks stemming from the Nvidia CUDA examples, the VSIPL-GPGPU-Library and the Rodinia benchmark suite. By applying MOLIS to the real-world benchmarks, up to 9% energy and 12% runtime can be saved.

Important Links:

Go Back