Social ranking criteria for pairwise gossiping in large-scale resource scheduling

Abstract

The concept of “online presence” has long been available only to people with enough technical background to complete a set of tasks with gory details. Thus, actual utilization of large-scale networks, such as grids and clouds, has not been realized until recently. With the advances in technology in multiple areas, such as multi-core CPUs, low-power energy-efficient FPGAs, virtualization, service-oriented architectures and web services, and autonomic computing, there has been an area of opportunity for the not-so-technologically advanced masses to actually take part in large-scale computing. Social networks are important to large-scale networking because they close one of the fundamental gaps: the trust between autonomous entities, which usually do not have a relationship history, or a ranking mechanism. One other common problem in large-scale networking is resource matchmaking, finding the right set of resource providers for a set of requesters, and vice versa. Traditional approaches to resource matchmaking use centralized repositories, which at the minimum does not scale well, among other issues. In this study, we propose adaptive pairwise gossiping protocols to take feedback from the system, based on existing basic social relationships, and trust levels between autonomous entities in the network. In addition to the ranking criteria we previously employed while selecting which nodes to gossip to, such as execution history, average distance, freshness of information, we also propose employing several social ranking criteria: overall popularity, trusted execution history, and social distance. By simulation, we show that (i) these social ranking criteria can be mapped to traditional ranking criteria in large-scale resource matchmaking, and (ii) the social ranking criteria perform comparably, based on several performance metrics. Moreover, we have a prototype social networking application that can incorporate such ranking criteria. We are still in the implementation phase, in which we are working on particular methodologies to measure and compare the performances of the two similar sets of ranking criteria in two different domains.