A growing trend in the development and deployent of grid computing systems is decentralization. Decentralizing these systems helps make the ore scalable and robust, but poses several challenges. In this paper we address one such proble - that of locating computing resources meeting specified requirements in a large scale heterogenous system. The heterogeneous and dynaic nature, coupled with the ultiple occurrences of these resources, makes the problem distinct from traditional data location problems found in the context of content-sharing systems. We propose GUARD (Gossip Used for Autonomous Resource Detection), a protocol that uses gossiping between neighbors to propagate the current knowledge of distances from available resources. GUARD is autonomous (all decisions are made locally, using knowledge based only on interaction with imediate neighbors) and does not make any assumptions about the underlying network topology. Our siulations show GUARD is ore efficient than other techniques such as random routing, history-based routing and frequency-based routing that have been used for similar purposes. We also show how GUARD can be odified to locateultiple categories of resources meeting multiple criteria.