that for 100% Normal traffic, behaviour is similar to normal ORWAR, including the utility-per-bit schedul- ing of bundles. As we now introduce different bundlesizes, we may not make optimal use of the available contact window anymore. Previously our bundles had a constant size of 5 bytes; with a transmission ca- pacity of either 10 or 20 bundles/sec, this means there was no risk of incomplete bundle transmissions or idle time as no bundles would fit in the remaining time. Furthermore, in this situation the estimation of contact window actually matters as it will determine whether bundles have to wait or will be trans- mitted. In these experiments we therefore analyse the difference in performance of ORWAR with and without retention classes in a predictable environment. Using the results from this analysis, we attempt to provide some indication of an optimum distribution of retention classes on the network for use in the realistic mobility model.
In the stationary model, the contact window is ∞. The scheduling of which bundles to transmit is therefore similar to the situation where bundlesize is equal for all bundles, though bundles are ordered by utility-per-bit rather than only their transmission priority. The results of this scheduling can be seen in Figure 6.28 for the three different retention classes.
We notice that the behaviour is similar to the case where we used a constant bundlesize. The main differences we see here is that the average loss is lower and no transmission priority suffers from 100% loss. This lack of 100% loss is caused by the fact that even low priority bundles may be transmitted if its size is small enough, rather than expiring or getting replaced by a higher priority and/or more recent bundle. Still, these lower priority bundles will still suffer from loss at a lowercthan higher priorities as the range of utility-per-bit values a low priority bundle can attain is lower than that of a high-priority bundle. Therefore, the majority of these low-priority bundles will still end up at the rear of the buffer.
The delay however shows different behaviour when adding different bundlesizes. This is most sig- nificant for Normal bundles, where the delay is not only significantly reduced but also shows completely different behaviour, evidenced in Figure 6.29. This is easily explained by the fact that lower priority bundles do not suffer from starvation by definition. Even low priority bundles may be transmitted with a low delay if their size is small enough. When using equal bundlesizes, the delay of the lowest priority would be maximal, while delay of higher priorities was kept to a minimum, until its loss reached 100%. At this point, the loss of the next highest priority would increase along with its delay. Now there still is a
(a) Normal
(b) Never Delete (c) Keep Most Recent
Figure 6.29: Mean delay in ORWAR for different retention classes, using the stationary mobility model
difference in delay between low and high priorities, yet the difference is not that big anymore, with even high priority bundles able to suffer from delay at lower creation rates.
When observing the loss of ORWAR using the linear mobility model, we again notice that the be- haviour is very similar to the behaviour when using constant bundlesizes. Again, similar to the stationary model, we see that there is no 100% loss for any transmission priority. Another difference compared to the constant bundlesize scenario we notice, is that loss occurs for Priority 4 bundles at a much lower creation rate. We already stated that due to nodeB’s mobility, loss occurs at a lower creation rate com- pared to the stationary model as the buffer fills up. Due to the varying bundlesizes, Priority 4 bundles may still end up in the rear of the buffer if their size is large enough. These bundles will then be replaced by bundles with a more favourable utility-per-bit until bundles can be transmitted to nodeBagain.
Furthermore, when we compare the loss of Never Delete and Normal bundles compared to the scenario using constant bundlesizes, we observe a significant reduction in loss. For Keep Most Recent bundles however, the loss reduction is less significant. This is due to different bundlesizes not affecting the fact that only the bundles inserted just before nodeBleaves nodeA’s range have a chance at arriving at the destination.
Similarly, the delay when introducing varying bundlesizes differs very little from constant bundlesizes. The main factor for delay in the linear mobility model is the travelling time of nodeBbetween nodesA
andC. Transmissions from source to destination are still done in bursts, with nodeAfirst transmitting its buffer contents to nodeBafter which nodeBwill transmit these contents to nodeC. The average delay of each burst is equal, dependent on the round-trip time of nodeB. Only for Never Delete bundles we see a slight decrease in delay. This is likely caused by newly inserted bundles not necessarily being inserted at the back of the buffer based on their relative transmission priority, increasing their delay.
When comparing the linear and non-linear mobility model, shown in Figure 6.32, we actually notice little to no difference compared to the linear mobility model. We expected to see increased loss as ORWAR reverts to best-effort transmission due to incorrectly estimating the contact window. Due to this best-effort transmission, some bundles may suffer from loss as the contact window ends mid- transmission. This however does not seem to be the case in this scenario.
One possible explanation for this is we chose our bundlesize to be smaller than the transmission rate of the network. This would decrease the number of incomplete transmissions relative to all trans- missions due to a wrong estimation of contact window. For every transmission that may be interrupted, there is at least one other that was correctly transmitted. Increasing the bundlesize will also increase the risk of incomplete transmissions. In addition, the actual contact window with nodeBrelative to the
(a) Normal
(b) Never Delete (c) Keep Most Recent
Figure 6.30: Loss per transmission priority in ORWAR for different retention priorities using the linear mobility model
(a) Normal
(b) Never Delete (c) Keep Most Recent
Figure 6.31: Mean delay per transmission priority in ORWAR for different retention priorities using the linear mobility model
(a) Normal
(b) Never Delete (c) Keep Most Recent
Figure 6.32: Loss in ORWAR for different retention classes using the non-linear mobility model
transmission rate is large enough for a node to completely transmit its buffer contents during the contact window. The buffersize of all nodes is 500 bytes, furthermore we have a transmission capacity of 100 bytes/second. During one contact opportunity, we can transmit8∗100 = 800bytes, which is more than enough to transmit the complete buffer. Therefore there is barely any difference between best-effort mode and an accurate contact-window estimation. For reduced performance to occur, nodes Aor B
should not be able to transmit its complete buffer contents during one contact opportunity.