Review - High Availability, Scalable Storage, Dynamic Peer Networks: Pick Two

This paper argues on the practicality of having P2P systems that aim to
provide ALL three properties - "scalability, storage guarantees and
resilience to highly dynamic memberships" as infeasible. Their argument
is based on today's hardware trends where bandwidth is the limited
resource and the amount of bandwidth required to maintain such
guarantees is too high.

This paper does a good job at highlighting the main problem with the
types of guarantees these P2P networks aim to provide - enormous
cross-system bandwidth to move state and data around. It shows
analytically how the resources of a typical computer connected to
broadband or dial-up can be exhausted with just maintenance information,
let alone query traffic and actual download traffic. It then proposes a
few solutions as to how to make these P2P systems work well in practice,
e.g. restricted access, or a few nodes with high connectivity and very
low join/leave rates.

The paper based its assumptions on the state of today's technology.
However, I believe that these problems could be resolved. As we
discussed in class, the telcos already have the technology to deploy
high bandwidth connection at the last mile, but not the incentive to do
so. I also believe that the assumptions are very strict in that it
expects all data to remain in the system and transferred each time a
host joins or leaves. At a greater complexity cost, we can imagine a
system that tolerates some degradation in replication factor (here they
assumed k=20 with no good justification) if it can expect another host
to join according with some copies of the files. Also we can imagine
that hosts that join the network could potentially reuse the files it
had prior to leaving, thus saving transfer bandwidth.

The problem of a high churn rate in P2P systems seem to be a fundamental
problem in that it breaks a lot of things. Is this due to the general
mentality of people to switch off their computer, unreliability of
cable/DSL modems (or companies that reset the connections), or the fact
that hosts with dial-up connectivity were taken into account in the
study? If the latter condition is true, then obviously churn rate will
be very high!
Received on Mon Nov 14 2005 - 01:05:04 EST