Wireless Ad Hoc and Sensor Networks

122 Wireless Ad Hoc and Sensor NetworksThe above algorithm being end-to-end can be implemented at theingress node as a congestion control software agent. No internal informationfrom the network is necessary for the implementation. Using theacknowledgments packets from the destination, the congestion controlagent estimates the number of packets in the network, including those atthe destination, by treating the network as a buffer attached to the egressnode. Here, packets entering the ingress and egress nodes can provide anindication of the bottleneck in the network, whereas the packets in thebuffer at the egress node along with the desired buffer occupancy canprovide information about the available buffer space. This informationwill be communicated back as part of the acknowledgment. Using thisfeedback information, the rate at which the packets are allowed to enterthe network is determined and used in the subsequent time interval.If packets entering the network at the ingress node are higher than theones reaching the egress node, then a bottleneck is expected soon. Usingqueue length at the destination and number of packets in the network, theonset of congestion is estimated and a suitable rate to mitigate congestionis calculated. Subsequently, the rate of transmission of packets is adjustedso that it can either increase or decrease from the previous time instant.3.5.2 Overhead AnalysisIn the implementation, we use the three parameters — buffer occupancy,packet arrival, and packet service rates at the destination — as feedbackinputs. The feedback information will add some overhead to the networkflow, but actually, we can get the value of packet arrival rate at thedestination buffer, given the buffer occupancy and packet service rate atthe destination buffer (packet arrival rate is equal to packet service rateplus buffer occupancy over measured time interval). Therefore, we onlyneed two parameters. We normalize these two values as a percentage oftheir maximum values (desired buffer occupancy and peak packet rate,respectively). Consequently, we need only 7 b to denote each value (percentage).So the overhead is only 14 b for a measurement interval. Thoughthis is small compared to an acknowledgment, future work will involvefurther reduction based on analysis on the overhead bits.3.5.3 General Discussion on ImplementationIn the implementation, we need only 14 b for a measurement interval.This is considerably smaller as compared to an acknowledgment packet.Moreover, the internal information from the network is not needed makingit an end-to-end implementation. The proposed algorithm does notdepend upon number of nodes in the network because the network istreated as a buffer and the number of packets in the network is estimatedas the buffer occupancy. Therefore, the algorithm is scalable.