Broadband traffic characterization and performance estimation
S. Devadhar, Mário Gerla · 1999
Recent studies have shown that data traffic possesses Long Range Dependence (LRD); specifically such traffic has no notion of a natural burst length, there is no dominant time-scale, and the bursts occur at many different time scales. As such, this traffic exhibits a self-similar or fractal nature. Such traffic puts a heavy load on the switching equipment, especially in terms of the buffering requirements. This research focuses on combining several areas within this field to yield useful metrics and tools that can subsequently be used for performance estimation. This study1 uses actual traffic measurements recorded in BAGNet (Bay Area Gigabit Network) a prototype network and Pacific Bell's production Frame Relay network. We specify a detailed measurement methodology and cite issues of particular importance that must be addressed while collecting broadband traffic data. Particular attention is given to issues related to measurement methods for delay calculation. Because of the prohibitively large amounts of data collected, we chose only certain subsets using the load or utilization as our primary criterion. Methods developed on BAGNet ATM measurements are subsequently applied to the Frame Relay data sets. Original descriptions of the real Frame Relay date in terms of parameters that define long range dependence are provided. The Hurst parameter, which is a strong indicator of Long Range Dependence (LRD), for 0.5 < H < 1.0. Using several methods, we show that for almost all of the traffic traces, this value is found to lie between 0.7 and 0.95. We next perform a burst-period (alternately busy-period) analysis of data from a trunk port with high levels of multiplexing. We find that the burst-periods are heavy tailed, whereas the idle periods are close to the exponential distribution. Next, directly measurement traffic delay is compared to that expected from analytical models. In our exploration of the histogram measurements on finite buffer queues driven by recorded traffic traces, an accumulation of mass is observed at the edge of the buffer. We call this the tail-raising effect. The implications of this effect are explored. Specifically, we investigate with empirical data, whether a heavy tail in the queue process necessarily implies heavy losses. A new estimator is proposed that exploits the information from the raised tail. An efficient bin-size selection and bin placement methodology, Symmetric Logarithmic Binning (SLB) is explored that is suitable for real-time measurements. It simultaneously serves a triple purpose of reducing the histogram counters, and is also adapted to the shape of the histogram (queue density) curve on a finite buffer, as well as to the spatial regions of interest in studying such a histogram. The efficacy of this scheme is explored and shown to be of definite engineering interest. Finally, an engineering methodology for performance estimation is proposed that combines virtual buffers, histograms, the new estimator and SLB for on-line performance estimation, is proposed. (Abstract shortened by UMI.) 1This work was performed in cooperation with Pacific Bell and Bellcore.