Energy efficient computation is becoming a main focus for computing systems design. It has become a necessity to consider power aspects when dealing with large scale data centric and high performance applications. Designing a large computing center can now require access to natural free cooling and cheap electricity in order to be economically viable. It is forecasted that this trend will continue and will reach a critical point when we transition to exascale computing i.e. systems capable of performing 1018 operations per second, a thousand fold increase relative to the current petascale capabilities. A significant change in the way we do computation is required in order to support such an increase in performance. This change among other things will have to come from using more efficient computing devices as measured by performance per watt. It has been shown, both theoretically and in practice, that using heterogeneous systems i.e. systems using different types of computing devices can be more power efficient than homogenous computing systems. One of the main challenges in using heterogeneous systems is the difficulties in programming and maintaining systems with different types of computing devices, programming languages, development tools, etc. This work is an attempt at furthering our knowledge of designing and programming power-efficient heterogeneous systems using programming languages and frameworks that operate at abstract and hardware agnostic levels.
展开▼
