Sender-initiated reliable multicast protocols based on the use of positive acknowledgments (ACKs) can suffer performance degradation as the number of receivers increases. This degradation is due to the fact that the sender must bear much of the complexity associated with reliable data transfer (e.g., maintaining state information and timers for each of the receivers and responding to receivers' ACKs). A potential solution to this problem is to shift the burden of providing reliable data transfer to the receivers-thus resulting in receiver-initiated multicast error control protocols based on the use of negative acknowledgments (NAKs). We determine the maximum throughputs for generic sender-initiated and receiver-initiated protocols for two classes of applications: (1) one-many applications where one participant sends data to a set of receivers and (2) many-many applications where all participants simultaneously send and receive data to/from each other. We show that a receiver-initiated error control protocol which requires receivers to transmit NAKs point-to-point to the sender provides higher throughput than a sender-initiated counterpart for both classes of applications. We further demonstrate that, in the case of a one many application, replacing point-to-point transfer of NAKs with multicasting of NAKs coupled with a random backoff procedure provides a substantial additional increase in the throughput of a receiver-initiated error control protocol over a sender-initiated protocol. We also find, however, that such a modification leads to a throughput degradation in the case of many-many applications.

Sender-initiated reliable multicast protocols, based on the use of positive acknowledgments (ACKs), lead to an ACK implosion problem at the sender as the number of receivers increases. Briefly, the ACK implosion problem refers to the significant overhead incurred by the sending host due to the processing of ACKs from each receiver. A potential solution to this problem is to shift the burden of providing reliable data transfer to the receivers—thus resulting in a receiver-initiated multicast error control protocol based on the use of negative acknowledgments (NAKs). In this paper we determine the maximum throughputs of the sending and receiving hosts for generic sender-initiated and receiver-initiated protocols. We show that the receiver-initiated error control protocols provide substantially higher throughputs than their sender-initiated counterparts. We further demonstrate that the introduction of random delays prior to generating NAKs coupled with the multicasting of NAKs to all receivers has the potential for an additional substantial increase in the throughput of receiver-initiated error control protocols over sender-initiated protocols.

The behavior of window flow control schemes that adapt to changing network conditions is studied. A dynamic window scheme, which adjusts the window size based on explicit network congestion indicators, is proposed. The throughput performance of the adaptive policy under both steady-state and nonstationary conditions is studied. The superiority of this scheme over some previously proposed adaptive window schemes based on implicit congestion detection is demonstrated.>