<p>The computational landscape is littered with islands of disjoint resource providers including commercial Clouds, private Clouds, national Grids, institutional Grids, clusters, and data centers. These providers are independent and isolated due to a lack of communication and coordination, they are also often proprietary without standardised interfaces, protocols, or execution environments. The lack of standardisation and global transparency has the effect of binding consumers to individual providers. With the increasing ubiquity of computation providers there is an opportunity to create federated architectures that span both Grid and Cloud computing providers effectively creating a global computing infrastructure. In order to realise this vision, secure and scalable mechanisms to coordinate resource access are required. This thesis proposes a generic meta-scheduling architecture to facilitate federated resource allocation in which users can provision resources from a range of heterogeneous (service) providers. Efficient resource allocation is difficult in large scale distributed environments due to the inherent lack of centralised control. In a Grid model, local resource managers govern access to a pool of resources within a single administrative domain but have only a local view of the Grid and are unable to collaborate when allocating jobs. Meta-schedulers act at a higher level able to submit jobs to multiple resource managers, however they are most often deployed on a per-client basis and are therefore concerned with only their allocations, essentially competing against one another. In a federated environment the widespread adoption of utility computing models seen in commercial Cloud providers has re-motivated the need for economically aware meta-schedulers. Economies provide a way to represent the different goals and strategies that exist in a competitive distributed environment. The use of economic allocation principles effectively creates an open service market that provides efficient allocation and incentives for participation. The major contributions of this thesis are the architecture and prototype implementation of the DRIVE meta-scheduler. DRIVE is a Virtual Organisation (VO) based distributed economic metascheduler in which members of the VO collaboratively allocate services or resources. Providers joining the VO contribute obligation services to the VO. These contributed services are in effect membership “dues” and are used in the running of the VOs operations – for example allocation, advertising, and general management. DRIVE is independent from a particular class of provider (Service, Grid, or Cloud) or specific economic protocol. This independence enables allocation in federated environments composed of heterogeneous providers in vastly different scenarios. Protocol independence facilitates the use of arbitrary protocols based on specific requirements and infrastructural availability. For instance, within a single organisation where internal trust exists, users can achieve maximum allocation performance by choosing a simple economic protocol. In a global utility Grid no such trust exists. The same meta-scheduler architecture can be used with a secure protocol which ensures the allocation is carried out fairly in the absence of trust. DRIVE establishes contracts between participants as the result of allocation. A contract describes individual requirements and obligations of each party. A unique two stage contract negotiation protocol is used to minimise the effect of allocation latency. In addition due to the co-op nature of the architecture and the use of secure privacy preserving protocols, DRIVE can be deployed in a distributed environment without requiring large scale dedicated resources. This thesis presents several other contributions related to meta-scheduling and open service markets. To overcome the perceived performance limitations of economic systems four high utilisation strategies have been developed and evaluated. Each strategy is shown to improve occupancy, utilisation and profit using synthetic workloads based on a production Grid trace. The gRAVI service wrapping toolkit is presented to address the difficulty web enabling existing applications. The gRAVI toolkit has been extended for this thesis such that it creates economically aware (DRIVE-enabled) services that can be transparently traded in a DRIVE market without requiring developer input. The final contribution of this thesis is the definition and architecture of a Social Cloud – a dynamic Cloud computing infrastructure composed of virtualised resources contributed by members of a Social network. The Social Cloud prototype is based on DRIVE and highlights the ease in which dynamic DRIVE markets can be created and used in different domains.</p>
Resource Allocation in Large-Scale Wireless Control Systems with Graph Neural Networks