Scheduling of multi-class single-server queues under nontraditional performance measures

We consider a multi-class production system without setups where many job classes share a single server. The traditional performance measure used for scheduling these systems is that of mean throughput time (i.e., the time spent in the system). However, mean throughput time may not be the only measure of importance in real systems. In particular, throughput time variance and the outer percentiles of throughput time may be equally important. We present two heuristics for scheduling multi-class single-server queues that are based on heavy-traffic analysis and perform well with respect to these nontraditional measures in a wide variety of cases. An approximation is given for the throughput time distribution under both scheduling methods.