Operating a Network Link at 100%

Abstract

Abstract. Internet speed at the edge is increasing fast with the spread of fiberbased broadband technology. The appearance of bandwidth-consuming applications, such as peer-to-peer file sharing and video streaming, has made traffic growth a serious concern like never before. Network operators fear congestion at their links and try to keep them underutilized while no concrete report exists about performance degradation at highly utilized links until today. In this paper, we reveal the degree of performance degradation at a 100% utilized link using the packet-level traces collected at our campus network link. The link has been fully utilized during the peak hours for more than three years. We have found that per-flow loss rate at our border router is surprisingly low, but 30 ∼ 50 msec delay is added. The increase in delay results in overall RTT increase and degrades user satisfaction for domestic web flows. Comparison of two busy traces shows that the same 100% utilization can result in different amount of performance loss according to the traffic conditions. This paper stands as a good reference to the network administrators facing future congestion in their networks.

Modeling Upper Layer Reaction to QoS Degradation as a Congestion Avoidance Mechanism

“…With simultaneous Eqs. (5) and (6) mentioned in Sect. 4.2, we can quantitatively calculate the above ratios for all traffic types and calculate the difference among the traffic types.…”

“…By solving Eqs. (5) and (6), we can obtain the solution of L t and actualized traffic T t given D t . This solution is actualized as an equilibrium of the closed-loop of traffic increase → QoS degradation → traffic reduction.…”

“…Because Eqs. (5) and (6) of the ULCL model cannot be analytically solved, we solved them numerically with the semi-Newton method. As a comparison, we also plotted the actual packet-loss ratio with the estimated D t .…”

“…We analyzed simultaneous Eqs. (5) and (6) to derive the equilibrium-loss ratio and actualized traffic by varying γ as well as when different γ values are mixed. We used δ = 0.006 to calculate actualized traffic variance † .…”

Inferring Latent Traffic Demand Offered to an Overloaded Link with Modeling QoS-Degradation Effect

“…Statistical analysis such as auto/cross-correlation among traffic and QoS time series have also been investigated [11]. Several measurement results on overloaded links have been published [3], [12], [13]. For example, Luckie et al investigated RTT (round-trip time) and packet loss time series of an inter-AS link and found that the diurnal pattern of RTT and packet loss increases to about a 70 ms delay and a 2 ∼ 3% packet loss ratio [3].…”