review of understanding availability

This paper discusses the problem of availability in p2p systems. The goal
is to design high-available systems by characterizing host availability.
The authors propose several interesting problems. First, previous
measurements underestimate p2p host availability. This is because a host
may have multiple IP address in a p2p system. For example, a host leaves
the p2p system, and comes back with a different IP address due to DHCP and
NATs. Second, host availability varies with the time period over which
availability is calculated. The result shows that host availability
decreases over time. Third, the authors use conditional probability to
characterize the dependency between host pair. They prove that the
interdependence between pairs of hosts can be ignored. Last, they show
the size p2p overlay network stays constant over the measurement period
due to the host turnover. The host arrival rate is close to the departure
rate. The authors claim that these characteristics should be taken into
account for providing highly available p2p services.

However, the host availability model is simple here. More factors should
be considered. One possible factor is peer’s capacity of processing query
and maintaining objects. There is an underlying assumption in current p2p
system. That is, hosts have the same capacity and workload. This is not
true according to the study in Gnuttella. Very few hosts do the main work
in p2p networks. In other words, some hosts are important than others. If
an important host goes down, it affects the file availability of p2p
system greater than other peers. This factor should be taken into account
in the host availability function.

This paper shows that host availability decreases over time. This is
mainly because they keep probing the same set of peers over the
measurement period. However, the p2p network is dynamically changing. The
arrivals and departures of peers maintain a stable p2p network in terms of
size, but with different peers. With replication schemas used in p2p, the
file availability does not necessary decrease over time.
Received on Sun Nov 13 2005 - 23:00:16 EST