Information Technology Reference
In-Depth Information
A communication system is used for real-time intercommunication, path latency
will directly impact the quality of connections [3]. For this reason, real-time response
is required for a distributed communication system and minimizing routing delay is a
primary objective. Additionally, communication systems are deployed in accordance
with the geographic session distribution pattern, the whole network shall be organized
as a multilayer architecture with several regions. Therefore, a hierarchical overlay is
necessary for the distributed communication system. Previous work proposes hierar-
chical P2P algorithms [4] where each hierarchy consists of super-nodes for upper
hierarchy. Super-nodes take more responsibility acting as centralized index, makes it
impractical for load-balance. The Comb protocol proposed in this article is purely
peer-to-peer and protocols on each node are completely the same. The whole overlay
is organized as a two-layered architecture, a Comb node maintains information about
(
ON
other nodes for routing, resolves a lookup in no more than two hops. Mainly
two features distinguish our design from many other P2P lookup protocols.
1. Comb is purely distributed. The Comb overlay is divided into several domains,
two-layered architecture satisfies the requirement of geographic distribution pattern.
Meanwhile, Comb abandons super-node, nodes in Comb overlay have no distinctions.
This makes Comb a load-balanced network, avoids single point of failure and perfor-
mance bottleneck. On the other hand, a purely peer-to-peer network is much better for
scalability.
2. Comb is simple and stable. A Comb node requires about
)
ON
size routing
table for lookups resolve, but routing performance degrades gently when routing in-
formation is out of date. Comb guarantees correct routing (though slow) of lookups as
long as one piece in routing table is correct. This is important for a distributed system
to keep steady.
The rest of this paper is organized as follow: Section 2 compares Comb to relevant
DHTs. Section 3 presents the system model. Section 4 is the base Comb protocol, and
Section 5 evaluates Comb's performance through simulation and experiments. Final-
ly, we summarize our contributions in Section 6.
(
)
2
Related Work
The first generation of DHT algorithms adopt a completely flat structure, the whole
overlay is organized as a ring or other plain topology. Consistent hashing [5] is uti-
lized to assign keys to nodes and resources. Generally, a DHT overlay consists of N
nodes which share R resources (N << R), the key space is partitioned randomly by
participating nodes, and each node is in charge of the resources belongs to its key
space section. Designs of flat DHTs include Chord[6], CAN[7], Kademlia[8], Pa-
stry[9] and Tapestry[10].
Flat DHTs have certain advantages, such as structural stability and workload ba-
lancing. On the other hand, they are incapable of achieving latency guarantees for
queries and offering a hierarchical overlay network. That makes flat DHTs improper
in to utilize in a distributed communication system. Hierarchical DHTs (HDHTs) [11]
are the last generation of DHT designs, HDHT nodes are distributed into hierarchies;
