Software power modeling method at architecture level based on complex networks

- An architecture level software power modeling method based on complex networks is proposed in this paper.
- We model software systems as complex networks and assume that the relation between the network characteristics of software and its power consumption is nonlinear.
- Experiment results validate the rationality of our assumption and BP neural network as an approximation tool is effectively.
- Our model shows that software power consumption could be analyzed from high level, which has important meaning for low-power software design.

The architecture of software systems can be naturally represented in the form of complex networks, especially for object-oriented software. Software as a kind of artificial complex networks, where entities of the software are nodes and interactions between entities are edges. These interactions are data-flows, instruction-flows and control-flows of the software, and these flows driving hardware circuit is the internal cause of power consumption of software. In this paper, we model software systems as complex networks at architecture level, assuming that the relation between the network characteristics of software and its power consumption is nonlinear. Based on this assumption, we propose a software power modeling method at architecture level. The model first measures network characteristics of software and then fit the nonlinear relation between the network characteristics of software and its power consumption by BP neural network. Experimental results show that our model could accurately estimate power consumption of the software, and the error is less than 11.2% compared to the measured value, which indicates our assumption is reasonable and our model is effective.