Decreasing Network Power with on-off Links Informed by Scientific Applications

As the tapering off of Moore´s Law produces the need for more parallelism in high-performance applications, the parallel programming model becomes central to achieving and maintaining performance per watt on current and future computer architectures. In order for a programmer to exploit parallelism in a power-efficient way, a close interaction with an abstract machine model is necessary to remain cognizant of important power-performance tradeoffs. This work introduces Asynchronous Circuit Programming (ACP), a way of explicitly identifying application communication as circuits to allow the underlying hardware to better execute code, and for a programmer to pipeline the allocation of communication hardware resources with computation. We provide a case study in using a preliminary C++ implementation of ACP with multiple scientific applications that illustrates how hardware can make use of application knowledge, keeping up to 75% of links in a low-power state.