Decentralized orchestration of composite services results in a number of flow topologies which differ in their messaging patterns and the distribution of flow code amongst different partitions of a topology. Different performance metrics result for different topologies, and this is governed by the availability of resources, such as CPU cycles, memory, network bandwidth, and so on. A performance model is used to evaluate the performance of different topologies based on availability of resources, and dynamically direct client requests between different topologies according to prevailing conditions.