Asymptotic behavior of generalized processor sharing with long-tailed traffic sources

We analyze the asymptotic behavior of long-tailed traffic sources under the Generalized Processor Sharing (GPS) discipline. GPS-based scheduling algorithms, such as Weighted Fair Queueing, have emerged as an important mechanism for achieving differentiated quality-of-service in integrated-services networks. Under certain conditions, we prove that in an asymptotic sense an individual source with long-tailed traffic characteristics is effectively served at a {em constant rate, which may be interpreted as the maximum feasible average rate for that source to be stable. Thus, asymptotically, the source is only affected by the traffic characteristics of the other sources through their average rate. In particular, the source is essentially immune from excessive activity of sources with `heavier'-tailed traffic characteristics. This suggests that GPS-based scheduling algorithms provide an effective mechanism for extracting high multiplexing gains, while protecting individual connections.