Legion is a data-centric programming model and a runtime system for writing and executing high-performance computing applications on parallel, heterogeneous architectures. In this paper, we characterize the Legion programming and runtime system in terms of its power and energy efficiency. We profile resource utilization, power and energy consumption, and performance of Legion-based applications on heterogeneous machines. We compare the power and energy consumption of traditional MPI programs with that of Legion-based versions. Our experimental results show that Legion applications can save up to 89.1% of execution time and 90.8% of energy, compared with the MPI counterparts. By dispatching application tasks to GPGPU with Legion, we can significantly reduce the execution time and the benefit is more obvious when the problem size becomes larger. The results also show that the power and energy consumption for identifying dependencies, mapping, and dispatching tasks on a Legion system is significant, requiring special attention in application and system design. To the best of our knowledge, this is the first study on the power and energy characteristics of Legion runtime and applications. Results and findings from this work will lead to efficient resource scheduling to maximize system performance while operating within a power envelope.
展开▼
