An Evaluation of Shared Multicast Trees with Multiple Cores

Native multicast routing protocols have been built and deployed using two basic types of trees: singlesource, shortest-path trees and shared, core-based trees. Core-based multicast trees use less routing state compared to shortest-path trees, but generally have higher end-to-end delay and poor fault tolerance. In this paper we consider a new type of shared multicast structure that uses multiple, independent, simultaneously-active cores. Our design provides for low end-to-end delay, improved fault tolerance, and low source discovery delay, while balancing bandwidth cost and routing state. These results indicate that shared trees with multiple active cores are a viable alternative to shortest-path trees. The Internet’s multicast routing structure is still evolving [1]. Since its inception in 1992, the Multicast Backbone [7] — the multicast-capable subset of the Internet — has primarily consisted of DVMRP [9, 26], PIM [10], and MOSPF [21] routers, tied together with a complex set of interoperability rules and utilizing a flat routing topology. In recent years, network operators have introduced native multicast support, policy, and a hierarchical structure. The current near-term solution consists of domains running PIM-SM [13] internally, connected by MSDP [15] for interdomain reachability. Eventually, MSDP will be replaced by BGMP [20], although some engineers advocate using the single-source architecture proposed by Express [8]. At the heart of all of these multicast routing protocols are two basic types of trees: single-source, shortest-path trees and shared, core-based trees. In each case, a set of senders wants to deliver data to a set of members, known as the multicast group. With shortest-path trees, a separate tree is built for each source, using the least-cost paths between the source and the members. With a shared tree, one tree is built for the entire group and is shared among all the senders. Building an optimal cost shared tree is known as the Minimal Steiner Tree problem and is known to be NP complete [30]. Core-based trees are a simple, low-cost approximation for this problem; a single router is chosen as the core, and a shortest-path tree is built from the core to the members. Senders transmit data toward the core until it reaches the tree. Shared trees are considered an important part of the multicast routing architecture because only one routing table entry is needed for an entire group, instead of one per source. Hence, BGMP uses shared trees for interdomain multicast to conserve state within the Internet backbone. Shared trees are also useful for rendezvous within a group, since the set of senders may not be known ahead of time. PIM-SM uses a shared tree in this manner, then switches to shortest-path trees once a sender is active. Despite these advantages, core-based trees have a number of drawbacks relative to shortest-path trees. Foremost among these is that core-based trees on average impose a higher delay between a source and the group members [29]. This is because packets often must travel first to the core and then to the group members, and the core may not be along the shortest path to each member. In addition, the core is a single point of failure; although PIM-SM use a list of backup cores [14], members may experience significant additional delay when a core fails. Finally, using core-based trees may cause traffic concentration, in which some links in the network are much more heavily utilized than others [29, 5]. Surprisingly, very little research has been conducted to study the possibility of using multiple active cores to ameliorate these problems. The designers of both PIM and CBT [2] considered using multiple cores, but chose to use a single active core, with backups available on standby, early in the design stage. OCBT ∗This work was supported in part by the National Science Foundation under grants ANI-9977524 and NCR-9714680.