LOSS SE N SIT IV IT Y

Average UtHy of Happy and Unhappy HEFS 2/3 HEFS and 1/3 (LEFS of IÆFS) (n=36f1=95%.ftto^.9) Intra Ctaas Fatm eas for rtgfi Expectaibn Flows Using 6*9 6 w lh 2/3 HEFS and 1/3 (LEFS or IÆFS) (Æ FS .n= 12,rho=0.9)

0.5 16 550 600 FIFO Hmax - 0.95 V A

(a) Average Utility (b) Intra-Class Fairness

Fig. 6.2: G-t-98 HEFS Average Utilities and Intra-Class Fairness for Two Different Expectation Mixes.

The second point is that the loss unhappiness of the HEFS is unaffected by whether MEFS or LEFS are grouped with HEFS. This is true as long as the proportion of flows is kept the same (two-thirds HEFS to one-third MEFS or LEFS). This result suggests that G-l-98 has the ability to separate flows based on their expectation groupings and treats them accordingly. The implication is that the dissatisfaction of the HEFS is partly attributed to the competition for buffer space among themselves.

6.3.2 A verage U tilitie s and Intra-C lass Fairness

Figures 6.2(a) and 6.2(b) display the average utility and intra-class fairness of the G4-98 HEFS applicable to both emix scenarios. Unlike the case of loss unhappiness, the effects of the aff allocation are recognisable in Figure 6.2(a). It is not clear, however, which one of the a n allocations is better. In some cases, one could have a higher average utility for the satisfied HEFS while at the same time have a lower average utility for the dissatisfied HEFS. Examples of these situations can be seen during the periods between 570-590 seconds for cr// equal to 4 and 510-530 seconds for (7// equal to 2.

Consider the interval 570-590 seconds. Since both configurations of G4-98 have very similar loss unhappiness, we can assume the happiness of some of the satisfied

C H APTER 6. LOSS SE N SIT IV IT Y 90

Loss unhappiness HEFS w*h 2/3 HEFS and 1/3 LEFS (n=36^i«55%.mo=0.9) Loss unhappiness Æ F S w lh 2/3 HEFS and 1/3 MEFS (n=36,ct=9% jho=0.9)

0.8 0.6 I 0.8 0.6 I 0.2 550 600 650

(a) two-thirds HEFS, one-third LEFS (b) two-thirds HEFS, one-third MEFS

Fig. 6.3: VBU HEFS Loss Unhappiness for Two Different Expectation Mixes.

flows with (jjj equal to 4 has increased while it has decreased for the dissatisfied flows. It is even possible that some of the happier flows became less happy. This means that, during these intervals, there are more variations in individual utilities for this configuration than when a n was equal to 2. These variations should lead to more unfairness. This is confirmed by the results of the HEFS intra-class fairness shown in Figure 6.2(b). Although the differences are small, the a n allocation of 2 is more fair during these times than the allocation of 4.

6.4

V B U w ith Different E xp ectation M ixes

In Section 5.3 we investigated the performance of VBU when used in a scenario similar to bmix conditions and found that the overall happiness can be increased by allocating suitable utility thresholds for less demanding flows while protecting the more sensitive flows. We now evaluate the performance of this scheme in a situation where the proportions of expectation mixes are unequal and skewed towards more demanding flows.

C H APTER 6. LO SS S E N S IT IV IT Y 91 6.4.1 Loss U n h ap p in ess

Figure 6.3 shows the HEFS loss unhappiness of two VBU configurations with utility threshold levels 0.10 and 0.05 which we call VBU-0.10 and VBU-0.05 respectively. Although the VBU schemes are not able to make all flows happy for both the emix-201

(Figure 6.3(a)) and emix-210 (Figure 6.3(b)) scenarios, we observe th at by lowering the utility threshold from 0.10 to 0.05 the number of unhappy users is decreased. This result is consistent with what we found in Section 5.3.2. As expected, there are more dissatisfied flows when the emix-210 flows are used than with emix-201. This is because the MEFS required more packets to go through to maintain the utility threshold of 0.10 (or 0.05) as compared to the LEFS. The resulting increase in the demand for buffer space has in tu rn affected the HEFS. In contrast, the HEFS of G+98 maintained the level of loss unhappiness regardless of the scenario used. The G+98 scheme is able to achieve this consistency by employing rate guarantees. This ensures th a t flows, especially the HEFS, have access to their buffer space when they are within their threshold. In the VBU scheme, this feature is deployed for the MEFS and LEFS while the HEFS are allowed to get as much space as they can get.

6.4.2 A verage U tility and Intra-C lass Fairness

Unlike the HEFS average utilities for the G+98 configurations, the shape and be­ haviour of the 0.10 and 0.05 utility thresholds are quite dissimilar. If we look at the upper half of Figures 6.4(a) and 6.4(b) where the average utilities of the satisfied HEFS are displayed, it appears th at neither VBU-0.05 nor VBU-0.10 can be consid­ ered better. However, as noted in Section 4.4, the average utihty plots should be used in conjunction with other evaluation indices to get a clearer picture. If we consider the times when the satisfied average utihties of VBU-0.10 are higher than those for VBU-0.05, we will find th at the loss unhappiness for VBU-0.10 is generally higher (which means more flows are dissatisfied) than for VBU-0.05. Examples of this situar tion can be seen in Figure 6.4(a) at times 425, 445, 475, 525, 545, 555 and 585 seconds. At these times, a few HEFS in VBU-0.10 managed to get very happy which resulted