Per-flow delay performance in traffic aggregates

Class-based traffic treatment frameworks such as differentiated service (DiffServ) have been proposed to resolve the poor scalability problem in the flow-based approach. Although the performance is differentiated in a class-based basis, the performance seen by individual flows in the same class may differ from that seen by the class and has not been well understood. We investigate this issue by simulation in a single node under FIFO, static priority, waiting time priority, and weighted fair queueing scheduling schemes. Our results indicate that such performance discrepancy occurs especially when flows joining the same class are heterogeneous, which is not uncommon considering that the same type of applications can generate traffic having very different statistical behaviors such as video traffic with different activity levels, or voice traffic with different compression schemes. We found that per-flow delay statistics, including the average and the 99^th percentile delay, can be very different from the corresponding class delay statistics, depending on flow burstiness, overall traffic load, as well as the queue discipline. We also propose a solution to reduce the mean delay variance experienced by flows in the same class.