In bulk power systems, spectral analysis of poorly damped modes is commonly used for identifying local and inter-area oscillations. Conventionally, these oscillations are mitigated by carefully tuned decentralized controllers. In this paper, we depart from the modal perspective and employ non-modal tools to analyze and control inter-area oscillations. Our input-output analysis examines power spectral density and variance amplification of stochastically forced systems and offers new insights relative to modal approaches. To improve upon the limitations of conventional wide-area control strategies, we also study the problem of signal selection and optimal design of sparse and block-sparse wide-area controllers. In our approach, we preserve rotational symmetry of the power network by allowing only relative angle measurements in the distributed controllers. We examine performance tradeoffs and robustness of sparse and block-sparse control architectures by studying the NETS-NYPS test system, and show that optimal retuning of fully-decentralized controllers can effectively guard against local and inter-area oscillations.