Design and implementation of self-organizing infrastructure for large-scale distributed services

Abstract

As computing and network capability increases, large-scale distributed services becomes popular where nodes are widely spread over Internet and the number of participants may scale to millions of nodes. Owing to fully decentralized architecture without central servers, they can obtain the ad-vantages of high scalability and robustness. In contrast, since participating nodes are volatile, that is frequently turned off, complicated recovery mechanism is needed. Also, it is difficult to develop them because they requires all middleware functions. For mitigating these difficulties, I proposed common self-organizing infrastructure helping these services. Each infrastructure node is provided by and belongs to individual organization. The key point is that infrastructure nodes are self-organized into adaptive hierarchical ring (AHR) overlay. In AHR overlay, each infrastructure node belongs to a ring and inter-rings are formed for interaction between rings. With the change of environment like the number of nodes, AHR overlay is reconstructed by ring division and merging for restricting the round trip time of rings within a desired de-lay bound. AHR overlay not only helps self-organizing configuration but also provides information broadcast channel where fresh information of each organization are broadcasted. It enables fast service and resource discovery in local infrastructure node. The self-organizing infrastructure provides another common service, scheduling proxy, for improving the inefficient scheduling of distributed computing. A scheduling proxy enables adaptive-grained scheduling between a master and workers by scheduling work units between them. It reduces the turnaround time of work unit and the waste of CPU time by timeout without increasing the performance cost of contention at the master.