A Reliable Multicast Framework for Light-weight Sessions and Application Level Framing

Paper Review: A Reliable Multicast Framework for Light-weight Sessions
and Application Level Framing

The paper proposes a scheme for reliable multicast, specifically for
light-weight sessions. The paper immediately argues that a one-size
fits all approach to reliable multicast is not suited for multicast
applications, which vary widely in terms of requirements. Their
framework, Scalable Reliable Multicast (SRM), only provides the
minimal definition of reliable multicast with no attention to delivery
order. Their scheme builds on the basic IP protocol with no necessary
modifications.

The paper shows that the issues facing multicast cannot be solved
using the methods applied in reliable unicast. The dynamics of
multicast requires unique solutions. A whiteboard application, wb,
and its assumptions are used to illustrate the key aspects of SRM.

In their scheme, session messages are exchanged to update control
information. These messages are limited to approximately 5% of the
aggregate data bandwidth. On the occurrence of data loss, a host will
multicast a repair request. The request/repair timer algorithm is a
major aspect of SRM. Multiple hosts may detect the same losses and
attempt to handle the same repairs. The operation of these algorithms
is also dependent on the topology of the system. The paper
investigates some common topologies and their implications on the
operation of the loss recovery. In the simple case of the chain, for
example, only a single request and repair message is needed. The
number of requests increases, however, in the case of star and tree
topologies.

The authors describe a set of simulations used to illustrate the
performance of their scheme. SRM performed well in small random
trees, but performance degraded (high number of repairs per request)
in the case of large, bounded-degree, trees. The performance was
dependent on the distance between the nodes in the tree. Their
simulation results in general are claimed to provide acceptable
performance. However, there was no standard timer parameter which
worked well across all topologies. Therefore, an adaptive algorithm
is proposed for continually updating the timer parameter. The paper
introduces a further optimization, local recovery, in which repair
requests are multicast to a limited area. This increases the
scalability of the system. At the time of the paper, this area was
not fully developed.

The use of simulations and the resulting graphs in this paper support
many of the author?s claims. However, I would suggest that the
authors provide more analysis of the graphs and provide explanations
to strengthen their results section which is somewhat unclear. The
authors chose to not take a generic approach to reliable multicast.
Although, if possible, it is advantageous to provide generic
solutions, for the sake of performance it is at times necessary to
take an application-specific approach. The authors chose this route
and the resulting framework is, in my opinion, more likely to be
adopted. The layering approach is not applicable in applications that
are not loss-tolerant and this work attempts to provide a novel
framework for those applications requiring reliable delivery.
Finally, the local recovery aspect of their proposal has significant
value and the authors should focus on this aspect of the framework if
they further this work.

NOTE: was unable to open pgs 10 and 12 of the paper in adobe.

-- Nadeem Abji
Received on Mon Oct 30 2006 - 20:12:10 EST