In this paper, we explore performance optimization problems for real-time systems that have to rely on a fixed energy budget during an operation/mission. We adopt the weakly-hard realtime scheduling paradigm to ensure a predictable performance for all the tasks: Our aim is to minimize the number of dynamic failures (in terms of (m,k)-firm deadline constraints) while remaining within the energy budget. We prove that this problem is NP-Hard in the strong sense even for an ideal DVS architecture with continuous speed spectrum. We propose techniques to statically compute the speed of the CPU in order to meet the (m,k)-firm deadline constraints. We present on-line speed adjustment algorithms to exploit the slack time of skipped and completed jobs. Through extensive simulations, we show how the performance can be significantly improved by selectively dispatching jobs by considering their energy costs as well as their contribution to the system performance.