Algorithms for Reliable Peer-to-Peer Networks

Algorithms for Reliable Peer-to-Peer Networks Rita Hanna Wouhaybi Over the past several years, peer-to-peer systems have generated many headlines across several application domains. The increased popularity of these systems has led researchers to study their overall performance and their impact on the underlying Internet. The unanticipated growth in popularity of peer-to-peer systems has raised a number of significant problems. For example, network degradation can be observed as well as loss of connectivity between nodes in some cases, making the overlay application unusable. As a result many peer-to-peer systems can not offer sufficient reliability in support of their applications. This thesis addresses the problem of the lack of reliability in peer-to-peer networks, and proposes a number of algorithms that can provide reliability guarantees to peer-to-peer applications. Note that reliability in a peer-to-peer networking context is different from TCP type reliability. We define a reliable peer-to-peer as a network that is resilient to changes such as network dynamics, and can offer participating peers increased performance when possible. We make the following contributions to area of peer-to-peer reliability: • we propose an algorithm that creates resilient low-diameter topologies that guarantee an upper bound on delays among nodes; • we study parallel downloads in peer-to-peer networks and how they affect nodes by looking at their utilities and the overall performance of the network; and • we investigate network metrics relevant to peer-to-peer networks and their estimation using incomplete information. While we focus on latency and hop count as drivers for improving the performance of the peers, the proposed approach is more generally applicable to other network-wide metrics (e.g., bandwidth, loss). Our research methodology encompasses simulations and analytical analysis to understand the behavior and properties of the proposed systems, and substantial experimentation, as practical proof of concept of our ideas, using the PlanetLab platform. The common overarching theme of the thesis is the design of new resilient network algorithms capable of offering high-performance to peers and their applications. As more and more applications rely on underlying peer-to-peer topologies, the need for efficient and resilient infrastructure has become more pressing. A number of important classes of topologies have emerged over the last several years, all of which have various strengths and weaknesses. For example, the popular structured peer-to-peer topologies based on Distributed Hash Tables (DHTs) offer applications assured performance, but are not resilient to attacks and major disruptions that are likely in the overlay. In contrast, unstructured topologies where nodes create random connections among themselves on-the-fly, are resilient to attacks but can not offer performance assurances because they often create overlays with large diameters, making some nodes practically unreachable. In our first contribution, we propose Phenix, an algorithm for building resilient low-diameter peer-to-peer topologies that can resist different types of organized and targeted malicious behavior. Phenix leverages the strengths of these existing approaches without inheriting their weaknesses and is capable of building topologies of nodes that follow a power-law while being fully distributed requiring no central server, thus, eliminating the possibility of a single point of failure in the system. We present the design and evaluation of the algorithm and show through extensive analysis, simulation, and experimental results obtained from an implementation on the PlanetLab testbed that Phenix is robust to network dynamics such as bootstrapping mechanisms, joins/leaves, node failure and large-scale network attacks, while maintaining low overhead when implemented in an experimental network. A number of existing peer-to-peer systems such as Kazaa, Limewire and Overnet incorporate parallel downloads of files into their system design to improve the client’s download performance and to offer better resilience to the sudden departure or failure of server nodes in the network. Under such a regime, a requested object is divided into chunks and downloaded in parallel to the client using multiple serving nodes. The implementation of parallel downloads in existing systems is, however, limited and non-adaptive to system dynamics (e.g., bandwidth bottlenecks, server load), resulting in far from optimal download performance and higher signaling cost. In order to capture the selfish and competitive nature of peer nodes, we formulate the utilities of serving and client nodes, and show that selfish users in such a system have incentives to cheat, impacting the overall performance of nodes participating in the overlay. To address this challenge, we design a set of strategies that drive client and server nodes into situations where they have to be truthful when declaring their system resource needs. We propose a MinimumSignaling Maximum-Throughput (MSMT) Bayesian algorithm that strives to increase the observed throughput for a client node, while maintaining a low number of signaling messages. We evaluate the behavior of two variants of the base MSMT algorithm (called the Simple and General MSMT algorithms) under different network conditions and discuss the effects of the proposed strategies using simulations, as well as experiments from an implementation of the system on a medium-scale parallel download PlanetLab overlay. Our results show that our strategies and algorithms offer robust and improved throughput for downloading clients while benefiting from a real network implementation that significantly reduces the signaling overhead in comparison to existing parallel download-based peer-to-peer systems. Network architects and operators have used the knowledge about various network metrics such as latency, hop count, loss and bandwidth both for managing their networks and improving the performance of basic data delivery over the Internet. Overlay networks, grid networks, and p2p applications can also exploit similar knowledge to significantly boost performance. However, the size of the Internet makes that task of measuring these metrics immense, both in terms of infrastructure requirements as well as measurement traffic. Inference and estimation of network metrics based on partial measurements is a more scalable approach. In our third contribution, we propose a learning approach for scalable profiling and prediction of inter-node properties. Partial measurements are used to create signature-like profiles for the participating nodes. These signatures are then used as input to a trained Bayesian network module to estimate the different network properties. As a first instantiation of these learning based techniques, we have designed a system for inferring the number of hops and latency among nodes. Nodes measure their performance metrics to known landmarks. Using the obtained results, nodes proceed to create anonymous signature-like profiles. These profiles are then used by a Bayesian network estimator in order to provide nodes with estimates of the proximity metrics to other nodes in the network. We present our proposed system and performance results using real network measurements obtained from the PlanetLab platform. We also study the sensitivity of the system to different parameters including training sets, measurement overhead, and size of the network. Though the focus is on proximity metrics, our approach is general enough to be applied to infer other metrics of interest, potentially benefiting a wide range of applications.