ANALYSIS OF DELAY LATENCY OF THE ACTIVE RELIABLE MULTICAST PROTOCOLS Lakhdar Derdouri * , Congduc Pham ** , and Mohamed Benmohammed * * LIRE Laboratory, University of Mentouri, Route Ain el Bey, 25000 Constantine, Algeria derdouril@yahoo.fr , mo.esi.ben@gmail.com ** LIUPPA Laboratory, University of Pau et des Pays de l’Adour, 64013 Pau Cedex, France congduc.pham@univ-pau.fr ABSTRACT This paper quantifies the reliability gain of combining classes for reliable multicasting in lossy networks in which the active network approach is the most promising. We define the delay latency of recovery as performance metric for reliability. We then study the impact of multicast group size and loss probability on the performance of compared approaches. Our simulation results show that combining classes significantly reduces the delay latency in lossy networks compared to the receiver-initiated class. Interestingly, combining classes can outperform receiver-initiated class depending on the network size and loss probability. Keywords: Active Networks, Reliable Multicast, Sender- Initiated, Receiver-Initiated, Delay Latency 1. INTRODUCTION Providing reliable and efficient multicast networking services in lossy networks is extremely challenging due to number of packets that can be corrupted or lost. To improve the reliability in lossy networks, the active networks approach has been proposed for multicast trafic. The active networks approach provides a user driven customization of the infrastructure in which new computations are dynamically injected into the nodes [6]. The use of active networks approach in reliable multicast has proven to provide more efficient solutions to the scalability problem for a large number of receivers. In active reliable multicast protocols, the members of a multicast group are organized in a distributed control tree to overcome the well-known acknowledgment implosion problem of flat approaches, i.e., the overwhelming of the sender by a large number of positive (ACKs) or negative acknowledgments (NAKs). In addition, the concept of active network can solve the repair locality problem in an effective way by attributing the role of repair to the router close to a loss. Several active reliable multicast protocols have been proposed such as ARM (Active Reliable Multicast) [10], AER (Active Error Recovery) [8] and DyRAM (Dynamic Replier Active reliable Multicast) [14], AMRHy (Active Multicast Reliable Hybrid) [1]. AMRHy and DyRAM are two protocols that use active services within routers. Each of them adopts a different strategy to solve the scalability problems. DyRAM belongs to the receiver-initiated class where the responsibility of loss detection is attributed to receivers regardless of the link on which the losses occur. In contrast, by combining the receiver-initiated and sender-initiated classes, AMRHy distributes the responsibility of loss detection between the source and the receivers. In this hybrid approach, the source handles the losses occurring in the source link while the receivers take care for those occurring in the tail links, thus providing an efficient distribution of loss recovery burden. This paper analyses the delay latency of the above mentioned two protocols in the presence of spatially correlated loss. Our simulation results show that the approach combining classes (AMRHy) provides good scalability and low delays compared to that based on the receiver-initiated class (DyRAM). Interestingly, combining classes can outperform the receiver-initiated class depending on the network size and loss probability. The remainder of this paper is structured as follows: In Section 2 the existing works on analysis of reliable multicast protocols are reviewed. Section 3 presents the description of AMRHy and DyRAM protocols. Section 4 shows the network model and hypothesis. Section 5 presents the simulation results of the delay latency analysis. Conclusions and directions for future works are presented in Section 6. 2. BACKGROUND AND RELATED WORKS The first comparative analysis of sender-initiated and receiver-initiated reliable multicast protocols was done by Pingali et al. [17]. This analysis showed thatprotocols of the receiver-initiated class are far more scalable than protocols of the sender-initiated class because the maximum throughput of the latter class is dependent on the number of receivers, while it is not the case for protocols of the receiver-initiated class. Levine et al. [11] have extended this work to ring-based and tree-based approaches and showed that the hierarchical structure organization of the receiver set in a tree-based approach guarantees scalability and improves performance. They also demonstrate that protocols based on the receiver-initiated class can not prevent deadlock when they operate with finite memory. Another comparative analysis of sender-initiated and receiver-initiated classes was presented by Maihöfer