Cache tuning has been shown to achieve considerable energy savings and methods have also been proposed for tuning the cache for standalone embedded applications. However, with the increasing complexity of modern day embedded applications, RTOS based multitasking systems are fast becoming the norm. Therefore, there exists a need for techniques to tune the cache for multitasking systems. In this paper we present a framework for energy centric tuning of the instruction cache for embedded multitasking systems. Our framework is built upon a formal model for characterizing multitasking systems and is suitable for fast instruction cache tuning using loop profiling. We validate our proposed techniques by applying them to tune a predictor based filter cache hierarchy - a common solution for low power embedded systems. For all the multitasking programs tested, our techniques are able to successfully predict configurations that are optimal or near-optimal. The proposed methods are also able to achieve speed-ups of up to an order of magnitude compared to exhaustive design space exploration techniques.