Analysis of NonGaussian Nature of Network Traffic and its Implication on Network Performan.pdfVIP

Analysis of Non-Gaussian Nature of Network Traffic and its Implication on Network Performance Tatsuya Mori†, Ryoichi Kawahara‡ Shozo Naito† tatsuya@, {kawahara.ryoichi, naito.shozo}@lab.ntt.co.jp †NTT Information Sharing Platform Laboratories, NTT corporation ‡NTT Service Integration Laboratories, NTT corporation 2 0 0 Abstract Self-similarity in data networks has been stud- 2 ied in terms of its ubiquitous presence and useful- p We analyzed the non-Gaussian nature of network ness from various measurements and statistical e S traffic using some Internet traffic data. We found analyses and simulation studies. However, these 3 that (1) the non-Gaussian nature degrades net- are not sufficient from the viewpoint of evaluat- work performance, (2) it is caused by ‘greedy ing network performance. That is, in some cases ] flows’ that exist with non-negligible probability, network traffic with a higher degree of LRD could I N and (3) a large majority of ‘greedy flows’ are show better performance than that with a lower s. TCP flows having relatively small hop counts, one — the reverse of above findings. This is be- c which correspond to small round-trip times. We cause LRD reflects only the temporal structure [ conclude that in a network that has greedy flows of traffic variability and not its spatial structure, 1 with non-negligible probability, a traffic control- such as marginal distribution. Grossglauser et v 4 ling scheme or bandwidth design that considers al. showed, using a flui