Disclosed are methods, apparatus and computer programs for improving performance and resource utilization of cluster-based data processing systems. The number of persistent connections between first and second data processing units of a system is adjusted according to monitored performance of communications between the nodes. The monitored performance measure is the queuing delay between the nodes, and the number of connections is increased and decreased as changing load conditions increase and decrease the queuing delay. A method of establishing and closing persistent connections in response to a monitored queuing delay between a front-end gateway computer and each of a cluster of back-end servers is used to improve the performance of a cluster-based Web server or provision of Web services. Unwanted connections are closed to enable system resources to be allocated to other processes.