In the ubiquitous computing environment, we believe that P2P-based context-aware application utilizing sensing data is important. However, since sensors are not placed under the control of one specific entity, it is difficult to retrieve appropriate information from the distributed sensors. In addition, when the same queries are sent by different users, same processes for collecting sensing data are repeated. The repetitive processing causes a load increase of sensors and network. In this paper, we propose a P2P-based architecture for managing distributed sensor information, which treats sensing data based on those semantic. The proposed architecture is constructed of three layers, 'Raw Layer', 'Abstract Layer' and 'Logical Layer', which enables to generate necessary information from retrieved data semantically. In addition, we propose a method to create virtual sensors by real peer. The virtual sensor manages processed data as a cache of real sensing data. We implemented the proposed architecture and method on PIAX, which is a development platform for agent-based P2P system. Then we evaluated the proposed virtual sensor placement method by simulations. By the simulation results, we confirmed that proposed method achieve a context-aware application with reducing the amount of network traffic.