The Process Group Approach to Reliable Distributed Computing
Kenneth P. Birman

This paper describes a system that attempts to simplify the building of
correct distributed systems/applications by providing reliable and ordered
multi-cast messaging in the communications layer.

The main contribution is the ISIS system that provides a common
implementation of this communication layer and exposes interfaces for its
use.  This includes the ISIS solutions to group membership addressing, group
communication and synchronization.  The communication and synchronization
features provide two interesting ordered communication protocols: ABCAST and
CBCAST.  The ordering for the protocols are derived from the concepts of
'close synchrony' and 'virtual synchrony' and provide absolute and causal
ordering on top of a reliable communication protocol. 

Although the paper benefits from having actual systems that use the ISIS
system containing the principals discussed in the paper, it appears that
these "real-world" implementations are relied on too heavily to provide
support for the performance and scalability of the system.  The paper does
not provide any experiment based metrics to support either the performance
or scalability, nor does it analyze the behavior of the existing
implementations for indications of future scalability in face of
developments in the processor, storage, and network infrastructures as well
as the increase in scale of systems such as the brokerage system used as an
example.  With the amount of network activity required by the protocols,
questions regarding scaling factors were bound to arise.

Although not really research oriented, more explanation on the increased
"ease-of-development" provided by ISIS would have been useful.  The authors
claim that the ISIS system allows even non-experts to build correct
distributed systems.  Exploration and research on whether ISIS truly
provided this ease-of-implementation could also have been used to validate
whether the interfaces provided by ISIS were the correct ones or whether the
supporting infrastructure should have been implemented in some other layer
possibly providing better support for implementation of correct systems.

The goal of providing packaged solutions to ease the development of
distributed systems in still a good goal in modern systems, and as such the
ISIS system, with its solutions, provides a good starting point.  However,
the internet and many distributed systems have exploded in scale during the
last decade and so the performance and scalability details would definitely
need to be examined.  The applicability of ISIS in modern systems would also
depend on the WAN implementation which was mostly skimmed over in the paper.