Click here for full text:
Building Low-maintenance Expressways for P2P Systems
Xu, Zhichen; Zhang, Zheng
Keyword(s): No keywords available.
Abstract: Recent P2P systems, represented by Oceanstore and PAST, offer an administration-free and fault-tolerant storage utility. Nodes in these systems collectively contribute towards a storage space, in a self- organizing fashion. While elegant from a theoretical perspective, they can be improved in three important areas: (i) low maintenance cost; (ii) the ability to make discriminative use of the nodes in the system that has different capacity and resource constraints; (iii) the ability to adapt to the underlying network conditions and the applications' needs. In this paper, we explore "expressways" as an auxiliary mechanism to deliver high routing performance, leaving the baseline infrastructure concentrating on tasks such as efficient resource utilization and simple management. The basic ideas of our proposal are to divide the total space in the overlay network into topology areas of different spans; we then select candidates in these areas to collectively construct a balanced, self- organized expressway system. By archiving the relevant system information (e.g. physical coordinates of the nodes) as objects stored on the system itself, we can further tune the expressway dynamically to fit both physical network conditions and application needs. Using CAN as the target system, we show that expressways can easily boost its routing performance from O(N(superscript 1/d)) to O(ln N), while keeping CAN's low maintenance cost. Our simulation studies show that our techniques to tune the expressways toward network conditions reduce the routing latency to about 50% over the default "random strategy" while staying within 2-3 times of the optimal. Furthermore, our simple heuristics to adjust the expressways toward application behavior can further reduce the physical latency up to 19%. Our mechanisms are robust and flexible, irrespective to the physical distribution of the participating nodes.
Back to Index