Group communication is a powerful and easy-to-use abstraction for distributed applications. The Totem protocol is a popular and efficient implementation of group communication primitives. To use Totem in soft real-time environments, the distribution of message latencies is an important performance measure — in particular, when fault tolerance is required. We conducted an experimental study of these distributions using a number of different fault scenarios (communication faults, performance faults, crash faults) and additional load scenarios. Using these experiments, we give empirical distributions for message latencies in the Totem protocol. The relative effects of various fault scenarios are discussed, highlighting the importance of scheduling delays for any practical, real-time group communication system.