Figure 6.14 shows the time series of byte throughput for UNC 1 PM in the inbound direction, revealing a good match between original and replayed traces. Lossless replays with and without collapsed epochs are generally closer than lossy replays, which are often 10 to 20 Mbps below the original. However, lossless replays show large spikes (minutes 14 and 21) that are not found neither in the original trace nor in the lossy replays. The lossy replays are actually very close to the original in the neighborhood of these spikes (e.g., between minutes 20 and 28). Interestingly, the time series for Leipzig-II shown in Figure 6.14 did not reveal a significant difference between lossless and lossy replays. Finding an explanation for this phenomenon requires further analysis, but this plot certainly justifies our choice of comparing the original trace to lossless and lossy versions of its source-level trace replay. Without a lossy replay, we would be tempted to conclude from the artificial throughput spikes in lossless replay that our source-level model is not properly reproducing an end-point limitation that was present in the original environment. However, the lossy replay, by showing that adding losses eliminates this spikes, demonstrates that they are purely due a network-level parameter and not to a limitation of the a-b-t model. Once again, we are not naively advocating for incorporating open-loop losses into traffic generation experiments, but addressing a difficulty that significant loss can create when trying to understand how realistic our modeling of the traffic source is. Simply relying on a lossless replay can be misleading, as this example demonstrates.
As in the full replay case, the lossless collapsed-epochs replay shows two large spikes that are not present in the lossy collapsed-epochs replays. The general impression from the plot is that collapsing epochs moderately increases the burstiness of the replay. Note for example the larger spike in the minute 5, the spikes in minutes 36 and 44, and the large ditch in minute 29. The collapsed-epochs lossy replay is quite similar to the full lossy replays, but we find a few periods where the approximation of the original throughput is slightly worse. For example, the collapsed-epochs replay shows a drop of byte throughput in minute 40 that is not present in the full lossy replay.
Figure 6.15 reveals somewhat different lessons from the time series of byte throughput in the outbound direction of UNC 1 PM. Regarding the full replays shown in the left plot, we see that the lossless replay has only one significant spike above the original traffic. One reason behind this finding is that the much higher average byte throughput makes spikes due to a few connections far less significant in relative terms.
Both full replays are generally slightly below the byte throughput of the original trace. The reason is not completely clear, but it suggests that the replay has a somewhat lighter distribution of connection throughputs, which makes the aggregate throughput slightly lower. If the replay is continued beyond minute 60, we do observe connections that remain active for a few more minutes and transfer enough data to account for the difference between the time series. We examined the logs from the generator hosts and confirmed that no overload occurred during the experiments, so the cause seems to be some artificial limit on the throughputs of the connections in our replay. One cause could be the overestimation of quiet times discussed in Section 5.2.1. Another possible cause is that the replays did not take into account the specific MSS of each connection. Every connection was given the FreeBSD default value (1,460 bytes), which is the most common one on the Internet. However, it could be the case that a significant fraction of the segments were carried in TCP connections with a smaller MSS. These connections would then have higher control overhead, making their transferring of the same payload result in more bytes and therefore higher aggregate throughput. Given the small size of TCP headers, it is unlikely that the extra overhead would result in more than a few additional Mbps.
The results from the replays with collapsed epochs are similar, although we observe several additional spikes in the case of the lossless replay. The lossy replay does not show these spikes, but it is still below the original for most of the time series. Interestingly, it provides a closer approximation in some regions, such as between minutes 10 to 22. We can argue that this is an accidental improvement due to the artificially larger throughputs that a fraction of the connections achieves after their epochs are collapsed.
The analysis of the packet throughput in the inbound direction shown in Figure 6.16 reveals a number of interesting characteristics. Both lossless replays show substantial spikes above the original packet throughput. This is consistent with the similar finding for byte throughput. We also observe that collapsed-epochs replays generated a substantially smaller number of segments than full replays. As in the case of the analysis of the Leipzig-II replay shown in Figure 6.3, the lack of detailed source-level modeling in the collapsed-epochs replays makes traffic less realistic in terms of the aggregate packet throughput. In contrast, the lossy full replay shows an excellent match for most of the time series. This result is different from the Leipzig-II one, where the full replays achieved a good approximation, but were still below the original packet throughput. Adding per-connection losses had a very minor impact on the Leipzig-II packet throughput, but the effect is substantial in the UNC 1 PM replay, where we observe increments of up to 2,000 packets per second. This result demonstrates the effectiveness of our source-level modeling method, and also justifies our effort to incorporate losses in the replay in order to study the realism of our modeling approach.
Figure 6.17 examines packet throughput in the outbound direction. Unlike the inbound direction, adding losses does not have a substantial impact here, and the aggregate packet throughput remains below the original trace even for the lossy full replay. As discussed in Section 6.2.2, this could be due to some limitations of our data acquisition algorithm in terms of how well it infers source-level characteristics, or to the use of the default MSS for all connections. As in previous cases, collapsed-epochs replays generate a substantially lower number of packets than full replays, which are far closer to the original packet throughput.
The marginal distribution of byte throughput for the inbound direction of UNC 1 PM and its replays are shown in the Figure 6.18. The bodies of the distributions show that lossy replays provide a better approximation, although they are slightly heavier than the original. Interestingly, the analysis of the time series in Section 6.3.1 showed lower aggregate throughput from lossy replays, which seems inconsistent with the heavier bodies in the marginal distribution. The explanation is given by the plot of the tails of the marginals, which shows far lighter tails from the lossy replays. The way in which losses were incorporated in the experiments limited peak throughput substantially at the fine scales considered in the marginal plots. This is because the probability of artificial losses increases linearly with throughput, which is not generally true for real conditions. On the contrary, the lossless full replay reproduced the tail very accurately, demonstrating that the experimental environment and generation method are perfectly capable of reproducing the observed peak throughputs. It seems likely that further refinements in the implementation of per-connection losses, making them less open-loop, could make the tails closer to the original.
The marginal distributions in the outbound direction, which are shown in Figure 6.19, reveal a somewhat worse approximation. We can distinguish three regions in the plot of the bodies. For values below 175 KB, lossless replays are lighter than the original, while lossy ones are heavier. Above 175 KB, all replays are lighter, which shows that the finding of lower aggregate byte throughput in Section 6.3.1 is due to overall lower throughputs at fine scales (rather than only to lighter tails). In the region after 175 KB, we can also observe that lossy replays are heavier below 275 KB and lighter above that. The marginal distributions from the lossy replays are less concentrated around the mean value, and are therefore somewhat more bursty, which is consistent with the similar finding for Leipzig-II (see Section 6.2.3).
Regarding the tails, we observe that for probabilities below 0.00075, the tail of original marginal is substantially heavier than the tails of the replay marginals. For probabilities above that, the collapsed-epochs replays show a major change in the shape of the distributions, being far heavier than the original for the largest values. We did not encounter a similar phenomenon in the Leipzig-II replays, where lossy collapsed-epochs replays always had a lighter tail than the lossless full replay. The number of 10-millisecond bins with very high throughput is larger for collapsed-epochs replays than for the full replays. Note that this artifact is only visible by looking at the tails of the marginals, and not at their bodies or at the time series of byte throughput.
The marginal distributions of packet throughput for UNC 1 PM inbound are shown in Figure 6.20. As observed for Leipzig-II, and as we may expect from 6.16, collapsed-epochs replays result in bodies that are significantly lighter than the body of the original marginal. Full replays are far closer, being the lossy full replay an excellent approximation of the original distribution. Interestingly, the tails reveal a rather different picture. Below 350 Kpps, the lossy replays have lighter tails than the original, especially in the case of the lossy full replay. Lossless replays closely approximate the original tail. Above 350 Kpps, both full replays are lighter than the original, while the collapsed-epochs replays reproduce the probability of very high throughput bins accurately.
Figure 6.21 shows the marginal distributions of packet throughput in the opposite direction. All replays are lighter than the original, being the lossy full replay the closest one. The tails from the replays are also significantly lighter than the original tail. We also observe a similar change in the tail of the collapsed-epochs replays, which are very close to the original for the largest values.
The left plot of Figure 6.22 shows that the wavelet spectrum of the original byte throughput in the inbound direction is well approximated by both full replays for lower and medium octaves. The finding of this good match at the lower octaves differs from the result for the replay of the Leipzig-II trace, where this part of the wavelet spectrum was not so well approximated. The lossy replay shows less energy for octaves 8 and above, while there is a significant jump in the energy of the lossless replay for octaves 12 and above. In the right plot, the lossless collapsed-epochs replays shows substantially more energy for octaves above 4, while the lossy replay provides a better approximation.
For the outbound direction, the left plot of Figure 6.23 reveals a better approximation of the finest scales by the lossless full replay, while both full replays closely match the original spectrum at coarser scales. The right plot shows that both collapsed-epochs replays have less energy at the finest scales, with a rather sharp ditch for octaves 5 and 6 that was not present in the original. This ditch was far less pronounced in the full replays. Beyond the finest scales, the lossless collapsed-epochs replay is a poor match of the original, while the lossy one provides a close approximation. This high impact of losses in the collapsed-epochs replay, far larger than in the full replay case, suggests a significant interaction between loss and long-range dependence when traffic is not generated according to a detailed source-level model. In other words, endpoints that generate traffic according to less realistic models (without epochs) are artificially more aggressive than Internet sources. This makes them more sensitive to lossy environments, since losses can more sharply decrease their higher throughput. This can result in experiments that overestimate the impact of losses on performance.
The estimated Hurst parameters and their confidence intervals shown in Table 6.3 are somewhat surprising. In the inbound direction, the estimated Hurst parameter of the original trace is most closely approximated by the lossless replays. The lossy full replay is slightly lower, and the lossy collapsed-epochs replay is far lower. The same is true in the opposite direction, at least for the lossless replays. It is difficult to interpret the meaning of these estimates in the context of the previous results. On the one hand, we found large spikes in the time series of byte throughput that suggest substantially higher burstiness in the lossless replays. Additionally, the wavelet spectra in Figure 6.22 did not find better approximations from the lossless replays. Notice for example that the lossless collapsed-epochs replay is clearly the farthest from the original. On the other hand, the tails of the marginal distributions clearly favored the lossless replays, showing lighter tails for the lossy replays. We could argue that the different metrics refer to different measures of burstiness, and conclude that adding artificial losses (using our open-loop method) makes the lossy replays less realistic in terms of Hurst parameter estimates. However, this conclusion seems too simplistic, since it is in contradiction with the Leipzig-II results. Adding losses made the estimated Hurst parameters far closer in that case. Assuming that the observed differences between the estimated Hurst parameters are significant, the reason for these divergent conclusions regarding the impact of losses must necessarily lie in some fundamental difference in the nature of the two network links. The estimated Hurst parameters say little about the difference, since all of the estimates are similarly high (above 0.92).
As discussed in Chapter 4, the Leipzig-II trace is a good example of university traffic dominated by downloading behavior (i.e., inbound traffic is substantially higher than outbound traffic). In contrast, the UNC 1 PM trace is dominated by content downloaded from UNC servers (rather than downloads from UNC clients) due to the presence at UNC of a major Internet repository of software and content, ibiblio.org. This made traffic volume and number of connections far higher for UNC. Still, why would these differences make introducing losses beneficial in the Leipzig case and detrimental in the UNC case for the approximation of the original Hurst parameters? We can speculate that the rate-limiting mechanisms used by ibiblio.orgcreate unusual loss patterns that are poorly approximated by our open-loop losses, but we do not have any supporting evidence.
The lessons from the analysis of the scaling in the packet throughput series is quite similar. The plots in Figure 6.24 show reasonably close approximations of the original by all of the replays in the inbound direction, and somewhat worse ones in the outbound direction. The spectrum of the lossless full replay provides the closest approximation to the spectrum of the original in both directions. The spectrum of the lossless collapsed-epochs replay is clearly not as close, showing a higher slope for medium to coarse time scales. As in the case of byte throughput, lossy replays show less energy than the original trace, especially for the fine scales in the outbound direction. Note also the systematic ditch around octave 14 for all four spectra from lossy replays. This suggests some unexpected periodicities at the 1-minute scale. A similar ditch can be found in the outbound direction of the original time series in octave 13, and this ditch is not reproduced by the replays.
Regarding the estimated Hurst parameters and their confidence intervals, Table 6.4 shows different results for the two directions. The estimates for the inbound direction confirm the lossless full replay as an excellent approximation, but here the lossless collapsed-epochs replay is also very close to the original. Both lossy replays are well below the estimated Hurst parameter of the original time series, and outside its confidence interval. The estimates for the outbound direction show again an excellent approximation by the lossless full replay, but here the lossless collapsed-epochs replay is far higher than the original and well within the non-stationarity region. Lossy replays are substantially better in the outbound direction, with the lossy full replay matching the original estimate.
As in the case of Leipzig-II, the lossy full replay of UNC 1 PM achieved a perfect match of the original time series of active connections This is clear for both the entire range of the time series shown in the left plot of Figure 6.26 using 1-minute bins, and for the 20-minute region shown in the right plot using 1-second bins. This finer scale view shows several sharp spikes (minutes 24, 28, 30, 35 and 39) that the lossy full replay tracked accurately. The lossless replay has only a slightly lower number of active connections per second, showing similar spikes (but with a negative offset in the y-axis). Collapsed-epochs replays had a far smaller number of active connections. Also, they did not track the features of the original time series so well. Notice for example the absence of the minute 24 spike in the collapsed-epochs replays.
Doctoral Dissertation: Generation and Validation of Empirically-Derived TCP Application Workloads
© 2006 Félix Hernández-Campos