Tipos según el espesor del fruto

In this subsection I present two related efforts used to construct QoS multicast trees. One involves the a priori calculation of a limited set of intra-domain routes in order to provide a choice of paths used to construct a tree. The other approach is a follow-on approach to a previous version of my work known as Yet Another Multicast (YAM) QoS protocol. It adds new features to the YAM architecture in order to constrain the cost of the one-to-many join mechanism.

In the subsequent sections, I discuss the goals, architecture, and specification of the YAM protocol and its use of a one-to-many discovery mechanism. I also discuss implementation issues as well as problems that arise when constructing a QoS sensitive tree over broadcast Local Area Networks (LAN). Finally, I discuss techniques used to reduce the control message overhead cost in using a one-to-many join mechanism.

* Match OR Fail (MORF): Alternate Path Multicast Routing protocol

MORF can be considered a reaction-type of protocol [108]. Its primary purpose is to find an alternate path to graft a new branch if the default or current branch does not satisfy the requirements of an application. The MORF architecture is divided into two parts: a QoS Manager and a Route Construction Agent (RCA). The QoS manager is a separate entity that can be co-located with the receiver on a host, or it can be attached to a leaf router. It is responsible for informing an on-tree router that the current branch does not meet certain requirements and that an alternate path is needed to graft a new branch.

The RCA is the entity that goes through the process of finding an alternate path. The important aspect to note is that the RCA calculates localized static routes. Hence, if a multicast tree spans several domains, only the alternate path functionality within a domain is exploited. In this way, routing updates do not have to be distributed throughout each domain. Beyond the exploitation of potential localized multipath routing, RCA's that are scattered throughout the internetwork can be used as a third party proxy for constructing a new path between the receiver and the source. Specifically, if a bottleneck exists between the RCA near the receiver and the source, the RCA can then attempt to find a new path between itself and another RCA that is near the source. Paths between RCA's are calculated via the Dijkstra algorithm [108].

Explicit routes are used by the RCA to protect against possible loops with alternate routes. If the explicit route reaches an on-tree node whose upstream path does not match the explicit route, a failure notification is sent down and the RCA updates its table to reflect the difference. Otherwise a match notification is sent back to the RCA. This Match or Fail characteristic is the basis for the term MORF.

One of the advantages of MORF is that its design allows it to be adaptive in its application to either a source tree or shared tree protocol. This is because action is only taken if the current or default branch does not meet the requirements of the application. However, the success of MORF is dependent on the existence of either multi-path intra-domain routing protocols, or in a distributed set of Route Construction Agents. In addition, there still exists a need to advertise the presence of other RCAs and to calculate Dijkstra routes on known RCAs. This implies that the set of RCAs should not exceed a number that becomes unmanageable in terms of distributing information and calculating routes.

• QoSMIC: Quality of Service Multicast Internet protoCol

QoSMIC is a recent effort that attempts to improve upon an early version of the YAM protocol, which is presented in the next subsection [38]. As in the case of YAM, QoSMIC attempts to build shared trees using a one-to-many join mechanism. This type of join searches for on-tree nodes that send responses to the originator of the join. The originator then uses the response(s) as the basis for deciding which path is to be used to graft a new branch onto the tree.

One of the concerns attributed to using a one-to-many join is the impact it has on the network in terms of the number of control messages that are generated. One strategy to reduce this impact is to use an expanding ring search for nodes on the existing tree, as opposed to a single one-to-many join that is propagated with a large Time To Live (TTL) value. However, the authors of QoSMIC have shown that an expanding ring search can be quite costly (exponentially higher than QoSMIC as the average degree of a router increases linearly). Hence, they introduce a new entity known as the Manager Router (MR). This entity acts as an intermediary node which knows the identity of the nodes that comprise a given tree. When a new node attempts to join the tree, it generates a local expanding ring search (i.e., one that has a limited TTL value), and also sends a query to the MR. Upon reception of the query, the MR informs nodes on the tree to advertise itself to the node that wishes to join the tree.

While the use of the MR represents promising work in attempting to reduce the cost of the one- to-many join, there are a number of concerns given its current design. One concern involves the amount of potential state that needs to be maintained by this entity. In addition, given the nature of its importance, fault tolerance addressing the single point of failure becomes a critical issue. To address this issue, QoSMIC would need to support an additional server replication protocol if a single primary MR were used for each group. Another approach would be to designate multiple M R’s per group.

A more intrinsic concern about MRs involves their designation and advertisement. One of the objectives of QoSMIC is “Limited Impact of Pre-Configuration Decisions” [38]. In the context of previous work such as CBT, PIM, this objective has related to the problem in selecting cores and advertising their existence before a multicast session has come into existence. The a priori

advertisement of <group, core> mappings leads to a potential explosion of state stored in routers. QoSMIC, as well as YAM, address this issue by having the core be a product of group membership. Ironically, however, the use of an MR re-introduces the problem of a priori selection and advertisement. In [38], the authors indicate that all leaf routers in all stub domains know the existence of an MR for a multicast group. To reduce this impact, one can assume that a hierarchy of MRs that advertise a range of multicast group addresses would need to be in place for a such a system to scale. Of course, this adds more complexity in the design and implementation of QoSMIC. Finally, the MR introduces a number of control message exchanges that can contribute to the delay in grafting a new branch onto a tree. This exchange involves:

• Query message: Leaf router to MR

• Advertisement Command message: MR to tree node(s)

• Advertisement message: Node(s) to Leaf Router

• Join message: Leaf Router to Node

The last message above is the one that grafts the new branch onto the tree. The topological distance between all the above entities determines the actual delay in grafting a new branch.

Other than the MR, and given that QoSMIC is based on YAM, many of the issues and design goals of QoSMIC and YAM are the same. These shall be discussed in more detail in section 3.2, where I present the design of YAM.