This paper described the architecture, design and implementation of a new
peer-to-peer storage system, TotalRecall, that is capable of managing
availability through performing different levels of redundancy in each
condition. Actually, the framework facilitates the system administrators to
specify a policy comprised from methods for (1) Availability Prediction (2)
Redundacny Management and (3) Dynamic Repair.
The abstract concept of policy is a general enough notion which helps the
system address a wide-spectrum of storage system trade-offs driven by the
choice of the administrators. One common drawback in the authors' evaluation
is their simplistic assumption on the homogeneity of hosts and their
constant bandwidth which is not true in real world applications. Also, their
experiment comprises only 32 machines, results of which may not be the same
as large scale p2p systems.
Although the authors present a concrete policy which is performing well, one
can incorporate machine learning algorithms for helping the system adapt at
run-time and learn how to adapt based on the observed performance.
Received on Thu Nov 24 2005 - 11:09:14 EST
This archive was generated by hypermail 2.2.0 : Thu Nov 24 2005 - 11:09:14 EST