Proceso de Pedidos, Programación y Confirmación

The overall contribution of A is to study the application of thePCNarchitecture in MANETs. After presenting a simulation study, the paper concludes that with some key modifications / extensions, theIETF’sPCNarchitecture is applicable in MANETs, and that the proposed mechanisms indeed improves the networks ability

to avoid thrashing situations. The study shows that using the unmodifiedPCN

architecture as described in the relevant RFCs, results in a poor performance. After surveying the challenges related to using distributed admission control (with

PCNspecifically) inMANETs, the paper proposes extensions/modifications toPCN

which target theMANET challenges and improve the performance significantly.

The most important modification is to introduce probing. As long as there is traffic on an IEA, the egress sends periodic reports back to the ingress identical to the original CL edge behavior described in [12]. However, if an IEA has been idle for a longer interval, the egress stops sending feedback reports on that IEA. When the ingress observes this, it is forced to initiate a probing session if new admission requests arrives for the given egress. This results in a scheme which combines the periodic reports of original CL with probing, both reducing the amount of signaling, and preventing the empty IEA problem, which are identified in the paper as the two most important issues. With original CL, the periodic signaling is sent on all IEAs always. With the proposed probing scheme, the

3.1. Network level congestion control

Figure 3.3. OriginalPCNmetering and marking. Only outgoing packets are

subject to metering and marking. Classifier ensures that only PCNtraffic is fed

to the meter and marker.

amount of signaling instead increases with the number of admitted flows (upper limited by the number of IEAs), which in most cases will represent a significant reduction. The problem with empty IEAs is avoided, as the probing session reveals the current status of the path prior to taking the decision. The proposed scheme uses only a single probe packet, which gives a relatively small delay increase and a low degree of intrusiveness. Probe packets are subject to normal metering and marking through the core, and the marking state is fed back to the ingress by the egress. It may however be favorable to use more than one probe packet for two reasons; to increase the robustness of probing in environments where packet loss is caused by random noise, or to mimic the requesting flow’s characteristics. These topics are not investigated in the paper.

Paper A also proposes modifications to thePCNmetering & marking algo-

rithms described in [11] (threshold and excess markers). The original algorithms intended for wired networks meter only outgoing packets on a link (see figure 3.3). However, since it is now a shared channel that is being monitored, all packets on the channel must be metered, which includes both incoming and outgoing packets (see figure 3.4).

This metering and marking scheme does not take the into account the hidden path issue as explained in section 3.1.1, as the congestion status on the channel around neighbors along the path is unknown. Thus, the admittance of a flow may cause over-admission on nodes that are one-hop neighbors to the nodes along the requesting flow’s path. This is a weakness of all the three protocols described in the related work section 3.1.2 as well (DACME, PMAC and SWAN). In the paper, it is suggested the possibility to feed 802.11’s CTS (clear to send) signaling messages to the meter in order to assess the resource usage around 1

Figure 3.4. Modified PCN metering and marking. Metering is now done on all packets received on air. Promiscuous traffic refers to packets captured in promiscuous mode, not intended for this node. Additionally, marking is done when a packet reaches its final destination.

hop neighbors. This was not explored further, as it would cause an unwanted dependency on the link layer. An alternative would be to introduce more signaling between neighbor nodes.

Channel variations caused by mobility or background noise may cause link breaks and re-routing, potentially leading to over-admission on alternative paths. To some extent this can be handled by the flow termination mechanism. However, frequent terminations increase the degree of thrashing in the network, as the resources already spent by a terminated flow can to some extent be considered wasted, as discussed in section 1.2. Instead, it is important to have a more conservative admission threshold. In the paper, the admission limit was set to 0.7 Mbps on an 11 Mbps channel in a network with relatively low degree of mobility. As also stated in the paper, there are still topics to be investigated further

with respect to applyingPCNmechanism in the MANETenvironment:

• The metering & marking algorithms should also take hidden nodes into

account.

• Sensitivity analysis on the respectivePCN parameters,

• Load control of best effort traffic is also needed, due to the shared channel

environment.

• Comparative studies with admission control schemes dedicated forMANETs, in order to assess the difference in performance withPCN.

3.2. Transport level congestion control

3.2

Transport level congestion control