Social profiles for dynamically configurable agents in large scale cloud, grid, and heterogeneous infrastructures

Abstract

A vast amount of computing resources are available throughout the world today. These are distributed worldwide, and are heterogeneous in platform, origins, motivations, ownership and control. Many large computational challenges also exist, which if addressed, would require a great amount of computing resources, often too much to be addressed by one single type of computing resource. In attempting to assemble enough resources to address these large challenges, two difficulties present; interoperability, and the diversity found in the social and economic backgrounds that both the work and resources may have. Multiple distributed agents offer an interoperability solution by wrapping services which need to interact with each other. Agents can represent computing resources and work, modelling and executing the types of interactions needed to make resource allocation decisions. In an ever changing resource allocation environment, agents need to be easy to create, configure and reconfigure. Social Grid Agents (SGAs) are designed to address resources allocation in grid computing. Being able to operate and communicate in a distributed heterogeneous environment, they offer a solution for interoperability. SGAs interact with one another and enforce the work allocation policies of the resources they represent. This is done using the Boolean matching features of the Classad language. But matching of subtle and nuanced social and economic backgrounds presents a challenge to Boolean matching mechanism. To overcome this challenge, agents need a mechanism which can describe and differentiate between the social and economic differences among the resources and work which they are allocating. This thesis extends previous research in Social Grid Agents. It enhances existing SGA technology to become more scalable and configurable. It also proposes using textual descriptions in unstructured language to describe the social and economic outlooks that computing resources and workloads may have. To measure the similarity between these descriptions, search technology is integrated into the agents. These descriptions are called ‘social profiles’. This thesis provides novel mechanisms to improve scalability, configurability and programmability of SGAs. Serialisation and class loading techniques are employed to achieve dynamic configurability of ‘live’ agents. An extensible and versatile format for the social profiles is developed. This thesis also provides mechanisms to exchange social profiles using distributed search technology and serialisation. Embedded search technology in SGAs allows quantitative similarity measurements to be determined for social profiles, yielding information useful in resource allocation decisions. An extensive set of experimental case studies demonstrate the efficacy of these techniques.