Review - Understanding Availability

An integral aspect of systems design is anticipating failure and recovering
from it. Resource availability in a peer-to-peer system can be thought of as
the quality of being "present or available for use." Unfortunately,
availability in a peer-to-peer system is highly time varying and these
systems must be resilient to hosts that join and leave the network
constantly.

The implementation of a peer-to-peer protocol must take into consideration
transient software failures. That is, it must tolerate many cases where
hosts disappear and return in short intervals. This study shows that nodes
join and leave the network an average of 6.4 times per day. The protocol
must also resist nodes that leave without notification, and cases where
communication is lost for only a short time. Most importantly, availability
is a function of time, which adds to its complexity.

Bhagwan et al. present an important argument. They assert that, due to
dynamic address assignment, active node probing is more accurately performed
on node identifiers. However, it is unclear if their dataset is entirely
complete. Considering the size of Overnet, this study only considered a
small percentage of all nodes. An indication of the inaccuracy of this study
is evident in Figure 4, which depicts the number of available nodes
continuously decreasing over time. This result is not well explained.

Overall, this paper does a good job of depicting availability, but it is
clear that longer test runs are required. While it is important not to be
intrusive on the network, accurate trends are not completely defined in a 7
day trace. The authors might consider increasing the time between probes to
elevate pressure on the overlay.
Received on Mon Nov 14 2005 - 09:51:51 EST