1 2 3 4 5 6 7
0 20 40 60 80 100
Characteristic Path Length
Number of removed nodes in % Plain Kademlia
Kademlia with MBO
Figure 3.6:Characterisitc path length under perfect attack (500 nodes).
We consequently have synthesized optimal sample topologies for both cases and derived the target motif signatures as parameters for the MBO module.
Subsequently, simulation models of MBO both for CAN and Kademlia have been implemented to examine the performance of our approach. The evaluation results show that optimizing topologies locally by engaging motif signatures significantly improves the topology balance in both investigated systems, expressed through even key space distribution for CAN and uniform indegree distribution for Kademlia.
MBO induces negligible messaging overhead in both cases. Even more importantly, MBO is extremely scalable as no information on the network state has to be exchanged at any time. It is based entirely on local information and computations. Therefore, with respect to complexity, MBO outperforms any current methods for distributed topology control.
3.6.2 Outlook
Optimizing topologies based on motifs so far has only been tested for common target motif sig-natures. However, topologies that are optimal with respect to other objectives may be charac-terized by a set of significance profiles which deviate for different nodes. Especially in networks where the nodes are not allpeers, but rather play different roles within the underlying topology.
One such example are optimally DoS-resistant overlay streaming topologies [86].
It is the objective of the next Chapter to explore whether network motifs still can successfully be engaged in local decision rules, despite the diversity of the system participants.
3.6 Summary and Outlook 63
4 Resilient Peer-to-Peer Live-Streaming Using Motifs
In the following Chapter we show how one can engage network motifs not just as an online topology control mechanism, but also to provide substantial resilience to attacks and high level of privacy to the involved participants in P2P based live-streaming networks.
High robustness against churn and resilience towards adverse behavior are the key require-ments for reliable P2P live-streaming systems. Their highly inter-dependent nature, based on the cooperative service delivery between all peers, necessitates a systematic and structural re-silience. This is very challenging to assure because of the self-organization and decentralized control of such systems. Reliable services, and hence a structural resilience, still are a vital pre-requisite for any commercial deployment of P2P live-streaming systems or IPTV infrastructures.
We propose an entirely distributed motif-based topology optimization to this end. Concise comparisons show that it creates resilient topologies comparable to state of the art, yet causing significantly less, almost negligible overhead with respect to both computation and messaging.
Our new approach also has another property of immense importance: it neither gathers nor forwards any information about the underlying network. Thus, it renders actions by malicious parties, aimed at disturbing the network operation, almost impossible as no peer can determine her/his position nor the position of any other peer in the network.
Such privacy is not guaranteed by any other state of the art method for P2P live-streaming networks, which makes our method outstanding in its area.
4.1 Introduction
Creating resilient topologies, and thus achieving high robustness of P2P streaming systems, is one of the main challenges when designing a new approach for this comparably new class of content distribution schemes.
P2P streaming promises to greatly relieve server load and aid in supplying large audiences with broadband multimedia content [87].
Two different services, video on demand and live streaming, create two fundamentally differ-ent problems in this field. While contdiffer-ent can be distributed to some or all of the audience in advance in the first case [88, 89], it is delivered directly upon creation in the latter, making it much more difficult to guarantee satisfactory provision without deficits.
Introducing the cooperative delivery of P2P systems, in which the streaming packets are re-layed between the nodes, further complicates the task. Forwarding peers are not only less performant than dedicated servers, but they additionally exhibit a much less reliable character-istic with high churn of frequently joining/leaving or even failing parties. Still, each node relies on the correct service of all preceding nodes on the packet path from the source, since failures and delays are propagated from peer to peer.
65
Moreover, using the system to deliver controversial content, or simply considering any com-mercial deployment, makes the system a worthwhile target for parties with malicious intent.
However, service degradation or even disruption, be they caused by churn, failure, or DoS attack is unacceptable. Especially when the service is targeted at deployment in commercial scenarios.
Achieving robustness and resilience with P2P streaming systems hence is a challenging task.
They represent interdependent large complex networks, composed of highly heterogeneous nodes with extremely dynamic behavior. Resilience in these systems consequently has to be achieved by decreasing the interdependence between the nodes while exploiting their resources to increase scalability. For reasons of complicated attacks by malicious parties, this task must be achieved with minimum disclosure of the system state.
In general, topology adaptation can either be implemented in a central oracle or by means of cooperative, distributed optimizations. Central instances represent single points of failure and potential bottlenecks, which makes them unfavorable when implementing large scale dis-tributed systems. Disdis-tributed adaptation, on the other side, is traditionally based on broad information on the overall state of the system, which is costly to gather and may be misused by malicious parties.
An alternative approach for cooperatively adapting topologies is to use metrics based on lo-cal information only and by optimizing the lolo-cal state, indirectly to approximate the overall desirable network characteristics.
In summary, we see the need to provide a scalable topology adaptation for live P2P streaming systems that successfully creates topologies of very low interdependency while considering only a bare minimum of local information.
The best distributed method (with respect to resilience) is currently a cost-based method.
It creates highly resilient topologies by considering aggregated information on succeeding nodes [86]. However, the engaged cost metrics are computationally demanding and knowl-edge about parts of the topologies has to be gathered. That knowlknowl-edge may potentially be exploited by malicious parties.
Network motifs on the other hand represent a metric much simpler to calculate. They also have the ability to reveal the complex interplay between topology and dynamics, see Chapter 2, and are a very promising application in topology adaptation, see Chapter 3.
In this Chapter, we propose to harness their ability to reproduce complex characteristics of networks for the purpose of adapting highly robust and resilient live streaming topologies.
We make the following contributions: we (i) present an innovative approach using network motifs to build local decision rules for the self-optimization of the network; (ii) improve the resilience of the topology to be close to the resilience of optimal topologies; (iii) evaluating our design in simulations, illustrate the benefits of our approach compared to the state of the art.
In order to better understand the challenges upon live-streaming applications, we first give a short background.