The parallel client-server paradigm

We consider the problem of modeling the algorithmic complexity of dynamic parallel client-server systems (PCS). The proposed PCS model is characterized by two features that distinguish it from common client-server systems. In the PCS model, both the client and the server are written and executed as parallel programs that use shared memory for communication. This is in contrast to the message passing style used by the common TCP/IP socket, for example, or by Java's remote method invocation. Secondly, the client and the server are also "dynamic" programs capable of adapting to any change in the number of machines that are available as processing units. We mainly focus on the problem of implementing fast queues where many elements can be inserted and deleted with a constant overhead per element. We use the PCS model to analyze the fast queue algorithm as an asynchronous algorithm and show its optimality. A novel criterion for measuring the asynchronous complexity of PRAM algorithms is introduced.