CnC: A Dependence Programming Model

Abstract

Application tuning is the one of the major hurdles on the road to exascale computing. Tuning is often directed at a specific architecture or towards some specific tuning goal. As currently practiced, the tuning activity requires serious expertise in the application domain, target architecture and tuning goals. Keeping all these (sometimes conflicting) concerns in mind at the same time while developing a program is very difficult and error prone. Dependence programming is a class of dataflow programming in which both data and control flow are explicit, are distinguished and are given equal weight. This paper gives an overview of CnC, a dataflow dependence programming model, from the perspective of the needs of exascale computing and shows how CnC addresses those needs through a separation of concerns. The goal of CnC is to enable a process that is easier, less error-prone and more effective, by separating these concerns into independent activities. Rather than proposing yet another approach to tuning, CnC provides a way to specify the application in a way that a) hides details not relevant to tuning, b) includes as much detail as possible to support the analysis and tuning process, and c) does not assume any specific architecture, style of tuning or tuning goals. This results in a specification of the application that can be applied to any existing or future architectures and can use any existing or future tuning approach. At the same time it can be more efficient in human time spent on creating the application and more effective in finding the best tuning.