Large-Scale Deployment in P2P Experiments Using the JXTA Distributed Framework

The interesting properties of P2P systems (high availability despite peer volatility, support for heterogeneous architectures, high scalability, etc.) make them attractive for distributed computing. However, conducting large-scale experiments with these systems arises as a major challenge. Simulation allows only to partially model the behavior of P2P prototypes. Experiments on real testbeds encounter serious difficulty with large-scale deployment and control of peers. This paper shows how an optimized version of the JXTA Distributed Framework (JDF) can help deploying, configuring and controlling P2P experiments. We report on our experience in the context of our JUXMEM JXTA-based grid data sharing service for various configurations. 1 How to test P2P systems at a large scale? The scientific distributed systems community has recently shown a growing interest in the Peer-to-Peer (aka P2P) model [1]. This interest is motivated by properties exhibited by P2P systems such as high availability despite peer volatility, support of heterogeneous architectures and, most importantly, high scalability. For example, the KaZaA network has shown to scale up to 4,500,000 users, an unreachable scale for distributed systems based on the traditional client-server model. However, the experimental validation phase remains a major challenge for designers and implementers of P2P systems. Validating such highly-scalable systems requires the use of large-scale experimentations, which is extremely difficult. Consider for instance popular P2P software, like Gnutella or KaZaA: workloads of these systems are not fully analyzed and modeled because the behavior of such systems cannot be precisely reproduced and tested [2]. Recently, P2P systems like CFS [3], PAST [4], Ivy [5] and OceanStore [6] based on smarter localization and routing schemes have been developed. However, most of the experiments published for these systems exhibit results obtained either by simulation, or by actual deployment on small testbeds, typically consisting of less than a few tens of physical nodes [7]. Even when larger scales are reached via emulation [8], no experimental methodology is discussed for automatic deployment and volatility control. For instance, failures are simulated by manually stopping the peers using the kill signal! There is thus a crucial need for infrastructures providing the ability to test P2P systems at a large scale. Several approaches have been considered so far.