1.1 Introduction
Recently, Peer-to-Peer (P2P) applications where no dedicated server is required, such as P2P media multicasting, P2P file sharing and P2P communicating software have become more and more popular. In P2P systems, peers with more resources (computation power, network bandwidth, disk space, etc.), often referred to as super nodes (SN), are essential for P2P services because many functions can be achieved with the assistance of a super node. For example, SNs can help in NAT traversal and application level routing support. Some P2P applications, such as Kazaa [1] and Skype [2] select a subset of the peers to handle several special tasks, such as searching files and routing support. Since extra tasks need to be performed by the supper nodes, load-balancing, how to evenly distribute the tasks among the super nodes, is also an important issue. In this thesis, we take the spatial locality into account and we present a super node selection strategy to select a super node close to the user.
1.2 Related Works
Super nodes are generally more powerful peers in terms of bandwidth, processing power and memory size. In Kazaa, super nodes are responsible for searching files and forwarding query messages to other super nodes. When a user wants to locate files, the user first sends query messages to a super node. In the IP communication platform developed by Internet Communications Lab, NCTU, we make use of super nodes to help users detecting NAT type and crossing NAT. If the
NAT type is a symmetric one [9], we need a super node to relay the media stream between the communicating pair. In these cases, if the super node is far from the communicating users, sending media streams through the super node may take up too much time and thus impair the performance. Since a P2P network is composed of a large number of hosts, naive strategies such as randomly selecting a super node or attempting to explore all nodes in P2P network seem not reasonable.
A well-known mechanism for selecting super node was presented in Gnutella protocol [3]. In Gnutella system, an ultrapeer (super node) acts as a proxy and relay query messages for other peers. Ultrapeers are used to reduce network traffic and speed up file searching. Since Gnutella is a decentralized P2P file-sharing system, any peer can determine if it is to become an ultrapeer by itself. A peer can be an ultrapeer only if following requirements are met: it is not firewalled, it runs on an OS that can handle large numbers of sockets, it has sufficient bandwidth, it has been up for at least a few hours and it has sufficient RAM and CPU speed. Although the strategy for selecting a super node in Gnutella system is simple and fast, Gnutella can not ensure that an ultrapeer is close to the peer it serves. Therefore, communicating delay may increase when an ultrapeer is far from the peer it serves.
Virginia Lo et al. [4] proposed a generic super node selection protocol in CAN [5], and Pastry [6]. Both are structured P2P networks and use a distributed hash table (DHT) to organize overlay structure and improve the searching ability. Virginia Lo presented a mechanism to store super node information in the DHT for fast and easy lookup. By utilizing the DHT, CAN and Pastry can enhance the searching performance. However, they don’t account for location information and it is hard to maintain the overlay structure.
PASS [7] is a novel P2P file-sharing system. PASS divides its overlay into multiple areas. Nodes are classified by using a geographical division such as a ZIP
code or an administrative network domain. Then, PASS selects a portion of high-capacity machines (super nodes) in each area for routing support. PASS shows that hierarchy structure and selecting super nodes by using location information can improves the file search performance. In PASS system, users may need to offer additional information so that PASS can assign them to a proper area.
In our thesis, we try to examine the location relationship between the user and the super node by measuring round-trip delay from the user to several super nodes.
1.3 Objectives
The objective of this thesis is to develop a P2P-like solution to establish and maintain a tree structure of super node group such that the each user is close to the super nodes of the same group and the overall distance of the entire tree structure is as minimum as possible. Therefore, it can reduce the transmission delay from the user to the super node. This mechanism can be applied to our previous research result, an IP communication platform that provides NAT traversal. We adopt the concept of hierarchy structure to reduce the network traffic in accomplishing our goals. We classify peers that can be super nodes into groups according to the transmission delay information. When a user needs a super node, he performs our algorithm to search a nearby group and choose one of the super nodes in this group. Since the super node may leave at any time and the media stream it serves needs to be relayed by another super node, fault tolerance is also a critical issue. We also propose a scheme to prevent a communication from being suspended due to the departure or failure of a super node.
1.4 Overview of This Thesis
The rest of this thesis is organized as follows. Chapter 2 describes the basic concepts and our super node selection algorithm. Chapter 3 presents the simulation results and analysis. Conclusions are given in Chapter 4