Normality of Ethernet Traffic at Large Time Scales

We contribute the quantitative descriptions of the large time scales for the Ethernet traffic to beGaussian.We focus on the normality property of the accumulated traffic data under different time scales. The investigation is carried out graphically by the quantilequantile (QQ) plots and numerically by statistical tests. The present results indicate that the larger the time scale, the more normal the Ethernet traffic.


Introduction
The experimental research of the internet traffic (traffic for short), including the Ethernet one, exhibits that fractional Gaussian noise (fGn) may be a model in the sense of unifractal see, for example, [1][2][3].This implies that traffic is Gaussian [4].However, non-Gaussian models, such as stable processes, were also reported; see, for example, [5][6][7][8].Therefore, the normality of traffic is an issue worth investigating.
Research described in [9,10] revealed a scaling phenomenon of traffic.Taking into account the scales of traffic, we say that whether a traffic trace is Gaussian or not relies on time scales.Paxson and Floyd [10] and Feldmann et al. [9] claimed that traffic is Gaussian at time scales larger than 1 second.That property was qualitatively further confirmed by [11].Note that real-traffic data used in [9,10] were recorded in 1980s and 1990s, which are publicly accessible [12].Thus, one second, as the critical time point, corresponds to the data in [12] and the infrastructure of the internet then.
Though the research exhibits that the statistics of traffic remain the same from the internet last century to the current years [13], the quantity of the critical time point, say one second, may be vague due to the development of high-speed networking.Therefore, when using the same data as those used in [1,3,9,10], we use the concept of packet count, that is, the number of packets within an interval, to represent the number of bytes of packets within an interval.
Let (()) be a sample record of traffic time series, where () ( = 0, 1, . ..) is the series of time stamps, indicating the time stamp of the th packet.The series (()) therefore represents the packet size of the th packet at time ().In this research, instead of using (()), we use () representing the packet size of the th packet.On an interval-by-interval basis, therefore, the accumulated traffic, denoted by (), is given by where  is the interval width, which also has the similar meaning of time scales.Thus, () stands for the accumulated bytes of arrival traffic in the th interval.The statistics of () may considerably differ when  is small (small time scale) or large (large time scale) [1,9,10].This research utilizes four real-traffic traces, listed in Table 1, which were measured on an Ethernet at the Bellcore Morristown Research and Engineering facility in 1989 [12].(the originally statistical properties described in the early literature, e.g., [1,3], turn to be ubiquitous in today's traffic, according to the research stated in [13].Thus, the traffic trace, BC-Aug89, which was measured in 1989, keeps its value in the description of traffic pattern today).
Figure 1 illustrates four series of real-traffic trace BC-Aug89.Note that the statistics of (()) is consistent with that of (), but we may obtain the time scale represented by  in (1), which is irrelevant of the networking speed.Let the interval width be  = 1024.Then, Figure 2 indicates () of BC-Aug89 for  = 1024.The paper aims at presenting the quantitatively minimum interval range for the accumulated Ethernet traffic traces to be Gaussian based on the accumulated bytes of the packets within an interval.
The remainder of this paper is organized as follows.In Section 2 we introduce briefly the commonly used normality tests and the idea of the QQ plot.The graphical and numerical results are presented in Section 3, and the discussion of the investigation results is followed in Section 4. Section 5 concludes the paper.

Statistical Investigation for Accumulated Traffic
In this section, we discuss the normality tests for the following null and alternative hypotheses: 0 : the data are sampled from a normal distribution; 1 : the data are not sampled from a normal distribution.
Many statistical tests have been proposed to find out whether a sample is drawn from a normal distribution or not [14], including the Shapiro-Wilk test, D' Agostino's  2 test, the Jarque-Bera test, the Anderson-Darling test, the Cramér-Von Mises criterion, the Lilliefors test, the Pearson's  2 test, and the Shapiro-Francia test.
The absence of exact solutions for the sampling distributions generated a large number of simulation studies exploring the power of these statistics.A convincing evidence from these studies is that convergence of the sampling distributions to asymptotic results was very slow.The paper [15] concludes that the Shapiro-Wilk test has the best power for a given significance, followed closely by Anderson-Darling test when comparing the Shapiro-Wilk, Kolmogorov-Smirnov, Lilliefors, and Anderson-Darling tests.On the other hand, some publications recommend the Jarque-Bera test [16,17].But it is not without weakness.It has low power for distributions with short tails.Therefore, we mainly consider three normality test methods listed in the following.

Shapiro-Wilk
Test.The Shapiro-Wilk test tests the null hypothesis that a sample (1), . . ., () came from a normally distributed population [18].The test statistic is where  () is the th order statistic;   is the sample mean;   is given by where  = ( 1 , . . .,   ); and  1 , . . .,   are the expected values of the order statistics of independent and identically distributed random variables sampled from the standard normal distribution, and  is the covariance matrix of those order statistics.It is worth mentioning that the Shapiro-Wilk test is restricted for the sample size greater than 3 and less than 5000.Normal QQ plot,  = 2 10  Normal QQ plot,  = 2 11   Normal QQ plot,  = 2 12  Normal QQ plot,  = 2 13  Normal QQ plot,  = 2 14   Normal QQ plot,  = 2 15  Normal QQ plot,  = 2 16  Normal QQ plot,  = 2 17

Anderson-Darling
Test.The Anderson-Darling test is a statistical test of whether a given sample of data is drawn from a given probability distribution [19,20].In its basic form, the test assumes that there are no parameters to be estimated in the distribution being tested, in which case the test and its set of critical values are distribution free.When applied to testing if a normal distribution adequately describes a set of data, it is one of the most powerful statistical tools for detecting most departures from normality [21,22], whereas the sample size needs to be greater than 7.

Jarque-Bera
Test.The Jarque-Bera test is a goodness-offit test of whether sample data have the skewness and kurtosis matching a normal distribution [23,24].The test statistic JB is defined as where ( If the data comes from a normal distribution, the JB statistic asymptotically has a  2 (2) distribution, so the statistic can be used to test the hypothesis that the data is from a normal distribution.For small samples, the chi-squared approximation is overly sensitive, often rejecting the null hypothesis when it is in fact true.Thus, JB test only applies to large sample size, at least 7 according to the finite sample study.
Besides statistical tests, we have another informal but powerful tool to assess the normality property of the series, that is, the normal probability plot.This graphical tool is often called the quantile-quantile plot (QQ plot) of the standardized data against the standard normal distribution.The correlation between the sample data and normal quantiles measures how well the data is modeled by a normal distribution.For normal data, the points plotted in the QQ plot should fall approximately on a straight line, indicating high positive correlation.
In order to obtain a more complete inference for the series' normality and to be more objective, we finally choose to take advantage of three popular normality tests, that is, the Shapiro-Wilk test, Anderson-Darling test, and Jarque-Bera test to verify the normality property in the application.Based on the software R, we mainly utilize the functions of the packages "fBasics" and "nortest" to realize the statistical tests.The -value of each test under the time scales  = 2  ,  = 9, . . ., 17 are presented in Tables 2, 3, 4, and 5.In particular, since the Anderson-Darling test requires the sample size greater than 7, there is no testing result for the time scale  = 2 17 .

Discussions
Graphically, from Figures 3, 4, 5 and 6, we have some findings listed below.
(i) Comparatively, the pAug.TL series asks for the relatively smallest time scale to be Gaussian among four series.(ii) The pAug.TL and pOct.TL data seem more likely to be normal than the other two series at each corresponding time scale.
(iii) It is not difficult to observe that the OctExt.TL and OctExt4.TL series exhibit the similar normality behaviors.However, only at quite large time scale, the theoretical normal quantile and the empirical quantile have the high positive correlation.Normal QQ plot,  = 2 10  Normal QQ plot,  = 2 11   Normal QQ plot,  = 2 12  Normal QQ plot,  = 2 13  Normal QQ plot,  = 2 14   Normal QQ plot,  = 2 15  Normal QQ plot,  = 2 16  Normal QQ plot,  = 2 17  (iv) The OctExt4.TL series seems to be even more strict on the time scale.It requires minimum time scale about 65536 to be Gaussian.
Numerically, as could be expected, the testing results given in Tables 2, 3, 4, and 5 provide the evidence that the larger the time scale, the more normal the accumulated traffic series ().Specifically, (i) it is straightforward to see that the normality behavior of pAug.TL data "surpasses" the others according to (ii) whereas, the pOct.TL and OctExt.TL series possess the comparable normality performance who need the time scale to be at least 32768 in order not to be rejected by the null hypothesis given the significance level  = 1%.(iii) for the OctExt4.TL series, in order not to reject the null, the time scale should be greater than 65536 given the significance level  = 1%.
The previous discussions are for the Ethernet traffic, but the methods may also be a reference for other types of time series, such as those in [25][26][27][28].

Conclusions
We have discussed the normality performance of the Ethernet traffic data under different time scales using several normality tests (Shapiro-Wilk test, Anderson-Darling test, and Jarque-Bera test).The graphical results by QQ-plot are consistent with the numerical results, which also provides the evidence for the quantitative results of the large time scales for the normality of the Ethernet traffic traces investigated.

Figure 3 :
Figure 3: QQ-plot of the pAug.TL data under different time scales.

Figure 4 :
Figure 4: QQ-Plot of the pOct.TL data under different time scales.

Figure 5 :
Figure 5: QQ-plot of the OctExt.TL data under different time scales.

Figure 6 :
Figure 6: QQ-Plot of the OctExt4.TL data under different time scales.

Table 1 :
Four traffic series.

Table 3 :
Normality test result for pOct.TL series.