Dragan Milic

QoS Enabled Multicast for Structured P2P Networks

In this paper we present a concept for providing QoS to multicast in structured P2P networks. We show on the example of Scribe / Pastry how to enforce QoS aware tree construction in a structured P2P network. We achieve this by modifying the ID assignment method of Pastry based on the QoS requirements of peers. As a result, the multicast tree holds the QoS (bandwidth) requirements on each of its end-to-end paths. We have evaluated the proposed concept by comparing default random Pastry ID assignment with our proposed method. The results of the evaluation show that using our method all end- to-end paths in the multicast tree fullfill the bandwidth QoS requirements, which is usually not the case for default Pastry. I. INTRODUCTION In this paper we present a concept to enable Quality of Service (QoS) for multicast in the structured Peer-to-Peer (P2P) network Pastry (1) and the Application Level Multicast (ALM) infrastructure Scribe (2) running on-top of Pastry. This is achieved by replacing the random ID assignment of Pastry with a QoS requirements aware mechanism. We use this concept to enable QoS for IP Multicast over ALM in the European project called EuQoS (3), which aims to support end-to-end QoS over heterogeneous networks. EuQoS only supports unicast QoS between end-systems on the network level. The remainder of the paper is structured as follows: In Section II we motivate the need for QoS enabled multicast Overlay Networks. In the same Section we present the use of such a network to provide QoS enabled IP Multicast services to end-systems in the EuQoS project. In Section III we present the ALM infrastructure Scribe and the P2P routing substrate Pastry. We present our proposal for modifying the Scribe / Pastry ID assignment to enforce the construction of QoS aware multicast trees in Section IV. We evaluate the QoS improve- ments by comparing default Scribe / Pastry ID assignment with our modifications in Section V. An outlook of future considerations and improvements can be found in Section VI. Finally, Section VII concludes this paper.

Supporting IP multicast streaming using overlay networks

In this paper we present our solution for providing IP Multicast on end systems in the Internet. The goal of the proposed solution is not to replace IP Multicast, but to provide an IP Multicast interface to applications on end systems in the current Internet environment, where IP Multicast is not available. Our solution, called Multicast Middleware, is a software, which is based on using Application Level Multicast (ALM) for transporting IP Multicast traffic. The use of the Multicast Middleware is transparent for applications on end systems, since our Multicast Middleware uses a virtual network interface to intercept native IP Multicast communication. In this paper we also present a performance evaluation of our Multicast Middleware. The results of this evaluation show that our Multicast Middleware is able to provide high bandwidth throughput to applications. This makes our Multicast Middleware a viable solution for supporting multimedia streaming services, etc.

Fisheye: Topology aware choice of peers for overlay networks

Constructing a topology aware overlay network is an open research topic. In this paper we propose a novel protocol for building overlay networks - a distributed fisheye view. Similar to round trip time (RTT) prediction approaches, we consider the end systems to be embedded in a virtual metric space. Unlike other approaches, we use only the distances (measured RTTs) to build an RTT proximity aware overlay network. Therefore, we are able to construct a fisheye view without performing the embedding. At the same time we are still able to guarantee the geographical diversity of the neighbors. Once built, the fisheye views on the end systems are continuously refined as information about new potential neighbors is available. This makes our overlay network adaptive to changes in the network topology. To evaluate our approach, we compared it with an existing topology aware overlay network construction approach - binning. We based this comparison on RTT measurements obtained using the King RTT measurement method and from the Planet Lab ¿all site ping¿ experiment. Our evaluation shows that the overlay network built using our approach outperforms binning in terms of relative RTT stretch. We also show that for increasing number of neighbors the performance of our approach converges towards an optimal solution.

Quality of Service for Peer-to-Peer Based Networked Virtual Environments

This paper describes how quality of service (QoS) enabled overlay multicast architectures using peer-to-peer (P2P) networks can enhance the experience of end-users in networked virtual environments (NVE). We show how IP multicast, which offers an easy to use API for implementing NVE but is not widely deployed, can be made available to end-users by bridging it transparently with P2P networks. We describe how different P2P and application level multicast (ALM) architectures can be extended with QoS mechanisms using our proposed OM-QoS (overlay multicast QoS) architecture. The presented approach allows users to experience QoS for NVE such as group-based multimedia broadcasting and distributed multiplayer games.

NetICE9: A stable landmark-less network positioning system

We propose NetICE9, a novel landmark-less method for embedding RTTs into virtual spaces. NetICE9 is inspired by VIVALDI, the most commonly used landmark-less approach. VIVALDI chooses its neighbors randomly and optimizes only towards one neighbor at a time. With NetICE9, we propose a solution to those drawbacks. NetICE9 significantly improves both the stability of the simulation and the precision of how RTTs are embedded into a virtual space. Our evaluation based on RTTs measured in the Internet show that NetICE9 significantly outperforms VIVALDI in terms of stability and precision of RTT prediction. NetICE9 improves RTT prediction also in comparison to GNP.

A performance comparison of native IP Multicast and IP Multicast tunneled through a Peer-to-Peer overlay network

We have developed the Multicast Middleware, a bridge between IP Multicast and a self-organizing Overlay Multicast infrastructure, in order to make IP Multicast available to the end user.We compare the performance of native IP Multicast and IP Multicast tunneled through a P2P Overlay Multicast network using the Multicast Middleware. We show that the achievable throughput can be quite high when using Overlay Multicast for data transport together with the IP Multicast API and that the latency introduced by processing (en- and decapsulation) and tunneling the captured packets can be negligible.

Guaranteed greedy routing in overlay networks

Greedy routing can be used in mobile ad-hoc networks as geographic routing protocol. This paper proposes to use greedy routing also in overlay networks by positioning overlay nodes into a multi-dimensional Euclidean space. Greedy routing can only be applied when a routing decision makes progress towards the final destination. Our proposed overlay network is built such that there will be always progress at each forwarding node. This is achieved by constructing at each node a so-called nearest neighbor convex set (NNCS). NNCSs can be used for various applications such as multicast routing, service discovery and Quality-of-Service routing. NNCS has been compared with Pastry, another topology-aware overlay network. NNCS has superior relative path stretches indicating the optimality of a path.

Multicast-Middleware der Universität Bern, Schweiz, resultierend aus EU FP6 IST IP „EuQoS“

Dragan Milic hat sein Informatikstudium 1998 an der TU Berlin angefangen und ist 1999 zur Universität Bremen gewechselt, wo er 2000 seinen Vordiplom erhalten hat. Sein Diplomstudium hat er 2004 an der Universität Bern abgeschlossen. Seit 2004 ist er Forschungsassistent und Doktorand am Institut für Informatik und angewandte Mathematik der Universität Bern in der Forschungsgruppe „Rechnernetze und Verteilte Systeme“.

Enhancing RTT prediction schemes using global function minimization

Numerous round trip time (RTT) prediction schemes use the least squares method to embed hosts in virtual euclidean spaces. The least squares method minimizes the residuals between measured data (measured RTTs) and their approximation (euclidean distances between the host position and fixed points, to which the distance was measured). This is achieved by minimizing an objective function, which is defined as a sum of square differences between measured distances to fixed points (landmarks) and euclidean distances to those landmarks in a virtual space. Since there is no direct way (closed form) for finding minima of the objective function, numerical function minimization must be used. In this paper we identify the problem of existence of multiple local minima of objective functions and their impact on resulting RTT predictions. To overcome this problem, we propose an algorithm for finding all local minima of the objective function. By finding all minima, we are able to identify the global minimum of the objective function, and thus ensure the optimal embedding of a host in the virtual space. To evaluate our algorithm we compare it with standard methods for function minimization using data collected by the Planet-Lab all-pings experiment.

Optimizing dimensionality and accelerating landmark positioning for coordinates based RTT predictions

In this paper we analyze the positioning of landmarks in coordinates-based Internet distance prediction approaches with focus on Global Network Positioning (GNP). We show that one of the major drawbacks of GNP is its computational overhead for a large number of landmarks and dimensions. In our work we identify two factors, which have a great impact on the computational overhead. The first one is being able to determine the optimal number of dimensions for embedding a given set of landmarks into a Euclidean space. The second factor is the selection of a good starting point for minimizing the total error of embedding. We propose an algorithm based on the simplex inequality (a generalized form of the triangle inequality) to extract the optimal number of dimensions based on distance measurements between landmarks. We also provide methods to compute a good starting point for the minimization problem and to reduce the number of variables involved in the minimization. We performed experiments with data obtained from the PlanetLab all-sites-pings experiment to verify the correctness and performance gains of our algorithm. The experimental results show that our enhancements to GNP landmark positioning are able to find the optimal number of dimensions for embedding the landmarks. These enhancements also accelerate the function minimization.