| Summary: | This work evaluates the performance of Peer-to-Peer storage systems in structured
Peer-to-Peer (P2P) networks under the impacts of a continuous process of nodes
joining and leaving the network (Churn). Based on the Distributed Hash Tables
(DHT), the peer-to-peer systems provide the means to store data among a large and
dynamic set of participating host nodes. We consider the fact that existing solutions
do not tolerate a high Churn rate or are not really scalable in terms of number of
stored data blocks. The considered performance metrics include number of data
blocks lost, bandwidth consumption, latencies and distance of matched lookups.
We have selected Pastry, Chord and Kademlia to evaluate the e ect of inopportune
connections/disconnections in Peer-to-Peer storage systems, because these selected
P2P networks possess distinctive characteristics.
Chord is one of the rst structured P2P networks that implements Distributed
Hash Tables (DHTs). Similar to Chord, Pastry is based on a ring structure, with the
identi er space forming the ring. However, Pastry uses a di erent algorithm than
Chord to select the overlay neighbors of a peer. Kademlia is a more recent structured
P2P network, with the XOR mechanism for improving distance calculation. DHT
deployments are characterized by Churn. But if the frequency of Churn is too high,
data blocks can be lost and lookup mechanism begin to incur delays. In architectures
that employ DHTs, the choice of algorithm for data replication and maintenance can
have a signi cant impact on the performance and reliability. PAST is a persistent
Peer-to-Peer storage utility, which replicates complete les on multiple nodes, and
uses Pastry for message routing and content location.
The hypothesis is that by enhancing the Churn tolerance through building a
really e cient replication and maintenance mechanisms, it will:
i) Operate better than a peer-to-peer storage system such as PAST especially in
replica placement strategy with a fewer data transfers.
ii) Resolve le lookups with a match that is closer to the source peer, thus con-
serving bandwidth.
Our research will involve a series of simulation studies using two network simulators
OverSim and OMNeT++. The main results are:
Our approach achieves a higher data availability in presence of Churn, than
the original PAST replication strategy;
For a Churn occuring every minute our strategy loses two times less blocks
than PAST;
Our replication strategy induces an average of twice less block transfers than
PAST.
|
|---|
