Hemera:SC:P2P-Ch

Robustness of large systems in presence of high churn (P2P-Ch)

Leaders: Pierre Sens (REGAL), Jean-Marc Vincent (MESCAL)

With rapid advances in networking technology and dramatic decreases in the cost of commodity computing components, large-scale distributed computing platforms with tens or hundreds of thousands of unreliable and heterogeneous hosts are common.

In these large infrastructures, the frequency of connections and disconnections (also called ”churn”) is a significant problem since the integrity of the network infrastructure completely depends on this important factor. So far, few studies focused on this issue. In the context of peer-to-peer systems, several simulation studies proved several weakness with respect to frequent joins and leaves (i.e. high churn). One of the main challenges is to maintain the connectivity of logical links in a scalable way. However, to our knowledge no real experiments on large testbed have been done. To study and observe large distributed systems (such as P2P overlays or Desktop Grids) in presence of realistic churn pattern is a crucial step in order to propose robust and scalable techniques.

Our objective in this challenge is threefold :

• Firstly, we plan to develop a formal model that would characterize the dynamicity of peer to peer networks. To validate our model, we will use traces available for p2p applications (Skype, Edonkey) and for distributed testbeds such as PlanetLab and Grid’5000.
• Secondly, based on this model characterizing churn, we will propose algorithms to resolve basic blocks of distributed systems such as leader election, consensus, resources allocation, and data storage. These algorithms will be based on a churn-resilient distributed overlay which dynamically adapts its structure to the current configuration.
• Finally, we will deploy these algorithms on the Grid’5000 testbed and develop some tools to inject dynamicity in Grid’5000 following our churn model. We will evaluate the churn resilience of different standard P2P overlays and test our new proposal. This final objective is to provide the first experience of churn injection in a large testbed distributed on more than 1000 different nodes distributed on at least 5 clusters of Grid’5000.