Section: New Results
Peer-to-peer systems
Participants : Pierre Sens [ correspondent ] , Jean-Michel Busca, Nicolas Hidalgo, Sergey Legtchenko, Sébastien Monnet, Gilles Muller, Corentin Travers, Véronique Simon, Mathieu Valéro.
Peer-to-peer storage
Since 2003, we develop Pastis [7] is a new completely decentralized multi-user read-write peer-to-peer file system. Pastis is based on the FreePastry Distributed Hash Table (DHT) of the Rice University. DHTs provide a means to build a completely decentralized, large-scale persistent storage service from the individual storage capacities contributed by each node of the peer-to-peer overlay
However, persistence can only be achieved if nodes are highly available, that is, if they stay most of the time connected to the overlay. Churn (i.e., nodes connecting and disconnecting from the overlay) in peer-to-peer networks is mainly due to the fact that users have total control on theirs computers, and thus may not see any benefit in keeping its peer-to-peer client running all the time. Since 2007, we study the effects of churn on Pastis, a DHT-based peer-to-peer file system. We evaluate the behavior of Pastis under churn, and investigate whether it can keep up with changes in the peer-to-peer overlay. We used a modified version of the PAST DHT to provide better support for mutable data and to improve tolerance to churn. Our replica regeneration protocol distinguishes between mutable blocks and immutable blocks to minimize the probability of data loss. Read-write quorums provide a good compromise to ensure replica consistency under the presence of node failures. Our experiments use Modelnet to emulate wide-area latencies and the asymmetric band- width of ADSL client links. The results show that Pastis preserves data consistency even at relatively high levels of churn.
However, when connection/disconnection frequency is too high in the system, data-blocks may be lost. This is true for most current DHT-based system's implementations. To avoid this problem, it is necessary to build really efficient replication and maintenance mechanisms. Since 2008 we study the effect of churn on an existing DHT-based P2P system namely PAST/Pastry. We have proposed RelaxDHT [43] , a churn-resilient peer-to-peer DHT. RelaxDHT proposes an enhanced replication strategy with relaxed placement constraints, avoiding useless data transfers and improving transfer parallelization. This new replication strategy is able to cut down by 2 the number of data-block losses compared to PAST DHT. We are now starting to study the use of erasure coding mechanisms along with replication within DHTs. Our goal is to propose hybrid mechanisms to find a good tradeoff among 1) churn-resilience, 2) maintenance cost, and 3) storage space.
Peer-to-peer overlay
Peer-to-peer overlays allow distributed applications to work in a wide-area, scalable, and fault-tolerant manner. However, most structured and unstructured overlays present in literature today are inflexible from the application viewpoint. In other words, the application has no control over the structure of the overlay itself. We proposed the concept of an application-malleable overlay, and the design of the first malleable overlay: MOve. In MOve, the communication characteristics of the distributed application using the overlay can influence the overlay's structure itself, with the twin goals of (1) optimizing the application performance by adapting the overlay, while also (2) retaining the scale and fault-tolerance of the overlay approach. The influence could either be explicitly specified by the application or implicitly gleaned by our algorithms. Besides neighbor list membership management, MOve also contains algorithms for resource discovery, update propagation, and churn-resistance. The emergent behavior of the implicit mechanisms used in MOve manifest in the following way: when application communication is low, most overlay links keep their default configuration; however, as application communication characteristics become more evident, the overlay gracefully adapts itself to the application.
We are considering a new class of target applications: massively multi-player online games (MMOG) such as virtual worlds. Within the context of a project funded by the LIP6, we have modeled a P2P distribution of such applications. Following this model, groups of object replicas are moving among peers while players evolve in the virtual world. For this kind of applications, it is important that the underlying overlay is flexible in order to remain adapted to the changing application structure. Since 2009, we are investigating an even more dynamic kind of overlay: a malleable overlay that anticipate application needs in order to adapt itself in advance. Thanks to this anticipation, the overlay is already operational and efficient when the application uses it.
Peer-to-peer publish-subscribe
Publish/Subscribe implemented on top of distributed R-trees (DR-trees) overlays offer efficient DHT-free communication primitives. We have then extend the distributed R-trees (DR-trees) in order to reduce event delivery latency in order to meet the requirements of massively distributed video games such that pertinent information is quickly distributed to all the interested parties without degrading the load of nodes neither increasing the number of noisy events. The enhanced structure performs better than the traditional distributed R-tree in terms of delivery latency. Additionally, it does not alter the performances related to the scalability, nor the load balancing of the tree, and neither the rate of false positives and negatives filtered by a node. The results of this work can be found in [57] which was also submitted to publication.
