Self-Similar Network Traffic and It's Control

点击翻译按钮获取中文摘要

Project Homepage: http://www.cs.purdue.edu/homes/park/nsf-ncr-9714707.html Abstract: Recent measurements of local area and wide area traffic have shown that network traffic is self-similar or fractal, i.e., the measured time series is bursty across several time scales. This phenomenon has been shown to be ubiquitous with potentially dire consequences to network performance due to the scale-invariant nature of burstiness. This project investigates effective traffic control algorithms, architectures, and mechanisms for facilitating efficient resource utilization while providing quality of service to real-time and best-effort traffic streams under self-similar traffic conditions. The project builds on recent work which characterizes the causal aspects of traffic self-similarity and shows that scale-invariant burstiness can be highly detrimental to network performance. In a nutshell, it is shown that self-similarity of traffic flow is an intrinsic property of networked client/server systems which communicate files or objects of very large size with nonnegligible frequency. The latter is known, technically, as the heavy-tailed property of file or object size distributions which is evidenced in actual distributed systems including UNIX file systems. On the performance side, it is shown that conjoint provision of low delay and high throughput is adversely affected by self-similarity. We follow a three-pronged method of attack for managing self-similar traffic, the first two based on dynamic control of traffic flow, and the third being of a structural nature where resources are architected and apportioned under direct control. The first method, predictive feedback control, is geared toward exploiting the long-range correlation structure present in self-similar traffic for congestion control purposes. We identify the on-set of concentrated periods of either high or low traffic activity - a distinguishing characteristic of scale-invaria nt burstiness - and adjust the mode of congestion control appropriately from conservative to aggressive. Being able to predict the onset of persistent congestion or idleness from fine time scale observations and utilizing this information effectively are two key issues. The second method, adaptive forward error correction, is directed at supporting multimedia traffic with real-time constraints without engaging in expensive resource reservations. Retransmission of lost information is not a viable strategy for such traffic classes since lengthy round-trip latencies encountered in wide area network environments render retransmitted information useless to time-constrained applications. We formulate and analyze an adaptive packet-level forward error correction mechanism called AFEC where the degree of redundancy is controlled as a function of network state, increasing the level of redundancy when the network is bad and decreasing it when the network is well-behaved. The control problem is nontrivial due to the fact that increasing the level of redundancy too much can backfire, further aggrevating congestion and causing the probability of timely information recovery at the receiver to decrease. We investigate the conditions under which adaptive forward error correction is effective when traffic is self-similar. We identify two regimes - packet loss dominated connections and delay-dominated connections - the former being most conducive to AFEC-based traffic control whereas the latter is more subtle, admitting only a limited range of useful redundancy. The third method, structural resource allocation, seeks to identify the relative utility of the two principal network resource types - bandwidth and buffer capacity - with respect to their curtailing effects on self-similarity. We conjecture a highly skewed trade-off relationship between bandwidth and buffer capacity in favor of bandwidth whose quantitative form becomes important when computing res ource reservations for traffic streams with stringent quality of service requirements. A bandwidth-over-buffer capacity policy also impacts the first two methods in terms of predictability and packet loss dominatedness/delay dominatedness. The three methods are expected to complement each other, each focusing on a different mode for quality of service provision under self-similar traffic conditions.