Reducing power consumption of communication networks is an important optimization goal in many application domains, ranging from large-scale simulation codes to embedded multi-media applications. Most of the prior efforts on network power optimization are hardware-based schemes. These schemes are predictive by definition as they control communication link status based on the observations made in the past. Since prediction may not be very accurate most of the time, these approaches can result in overheads in terms of both performance and power. This paper proposes a compiler-driven approach to communication link voltage management. In this approach, an optimizing compiler analyzes the application code and extracts the data communication pattern among parallel processors. This information along with network topology is used for identifying the link access patterns. These patterns and the inherent data dependence information of the underlying code help the compiler decide the optimum voltages/frequencies to be used for communication links at a given time frame. Our focus in this work is on loop-intensive codes which frequently appear in data intensive video and image processing. We exploit the regularity in data accesses of these codes to abstract out their inter-processor communication patterns, which in turn enable us select the most appropriate voltage/frequency level to employ for each communication link at any time.