Abstract: Centralized automatic service composition and so called non-trivial service discovery can retrieve compositions of services for a given request that contains particular inputs and outputs, and are now popular research topics in the area of service computing. Aiming at tackling the bottleneck and single-point-failure problems caused by centralized architecture, a distributed approach based on content-based distributed publish-subscribe technology, is proposed and in turn decentralized automatic service composition in large scale network is implemented. It is called process search in this paper. Content based distributed publish-subscribe system can deliver publication messages from publisher to interested subscribers hop by hop via a set of brokers by using content matching relationships among publications and subscriptions. It can offer a platform to support interoperability matching among service interfaces and to enable query routing in a distributed environment. In this paper, the service model is mapped to the message model of publish-subscribe system, the searching algorism is put forward based on content-based routing，and a prototype named PreSee is developed based on PADRES system. The stimulation experiments show that compared with centralized approach，this decentralized approach can reduce system latency by parallelizing searches among the distributed resources, thus improving efficiency of the whole system.