This paper proposes a computational model to predict communication skill of individual participants in online peer-to-peer dyadic interactions. Participants interact using a web-interface (skype-like) developed to record videos of two participants separately and stream real-time. This platform allows interviews and group discussions to be conducted at any place without the need of colocation of all participants. Towards studying communication skill in such a setting, we collected 72 dyadic interactions of participants discussing on general topics. Manual rating of these interaction videos have been obtained from two external observers. Multimodal features related to non-verbal behavior: prosody, speaking activity, visual and verbal behavior are automatically extracted. The experimental results show that the multimodal fusion of features with reliefF feature ranking perform best with an accuracy of 74%. Also, we automatically rate the overall discussion by extracting features from both participants. Our best results on automatic discussion rating achieves an accuracy of 83%. Further to dyadic interactions, we make a comparative study of the communication skill rating of the same participants giving an interface-based interview. Our study indicates that participants communicate well in interaction based scenarios.