Resource Co-allocation With QoS Constraints In Grid Computing

As resources are distributed and heterogeneous in large-scale grid environments, applications frequently require access to multiple resources acrossing different virtural organizations. Therefore, resource co-allocation is always a fundamental issue in grid computing, and co-allocation service has become a key service in grid infrastructure.With the development of grid technique, more and more grid systems have been deployed in commercial business, in which user's QoS requirements become more importent than before. These user's QoS requirements are often described as certain constraints, for instance, cost constraint, real-time constraint, reliability constraint, security constraint and etc. Many studies on performance evluation and analysis of grid systems indicate that conventional co-allocation service and policy are capable of meeting the requirements of traditional grid applications, however, it is difficult to satisfy those commercial grid applications. Therefore, co-allocation with user's constraints has become an important issue with therotical and pratical value.The challenges of co-alloction under user's constraints are focused as following:(1) it is difficult to coordinate the use's QoS requirements and the system performance at the same time; (2) existing co-allocation service can hardly guarantee all the user's constraints because of the dynamical and unpredicated workload of resources; (3) as the resource failures occur frequently in grid environment, it is difficult to prvide satisfied reliability guarantee for grid applications; (4) there is no suitable metric to quantatitively describe the dynamical capibility of grid resources.The studies in this thesis are based on the existing researches on performance evaluation and analysis in pratical grid systems. The statistic theory is applied to figure out the bottleneck of co-allocation under user's constraints. In order to model the working of grid resources, stochastic service theory is used to quantatitive analyze the runtime capability of resources. Meantime, the thesis proposes serval co-allocation policy and model to meet the various users'QoS constraints by using gaming theory, computing economic, graph teory, queueing theory and etc. The content and contributions are as following:(1) Advance co-reservation mechanism based on relaxed policyTo mitigate the negative effects brought about by advance reservation, a relaxed advance reservation policy is proposed, which allows accepting new reservation requests that overlap with existing reservations under certain conditions. Both the benefits and the risks of the proposed policy are presented theoretically. Experimental results show that the policy can achieve higher resource utilization rate and lower rejection rate comparing to conventional reservation policy and backfilling technique. In addition, the policy shows better adaptive when the grid systems are in presence of high reservation rate.(2) Resource co-allocation under cost constraintTo address the issue of resource co-allocation under the constraint of user's budgets in grid environments, a novel resource co-allocation model based on three-side gaming is proposed. In the proposed model, a new type of entities called Virtual Resource Retailer is introduced. In this way, typical resource co-allocation is modeled as a three-side gaming process, in which a cooperative gaming model and a non-cooperative gaming model are constructed to optimize the benefits of resource providers and costs of resource users through a flexible price mechanism. The validity and solution of the three-side model are presented theoretically. Extensive simulations are conducted to examine the effectiveness and performance of the proposed model in grid test-bed model by using real workload. Experimental results show that the three-side model outperforms conventional economy-based co-allocation model in term of resource benefits and resource utilization rate.(3) Resource co-allocation under real-time constraintA novel approach is proposed to evaluate the deadline-guarantee of a specific co-allocation scheme. Based on this approach, a multi-policy co-allocation model is proposed and implemented, which takes advantage of existing co-allocation policies to generate multiple co-allocation schemes and select the scheme that can provide optimal deadline-guarantee for applications. By using the workload obtained from the log of practical grid system, extensive simulations are conducted to investigate the performance of the multi-policy co-allocation model in multicluster computing grid test-bed. Experimental results show that the proposed model can provide co-allocation scheme with enhanced deadline-guarantee for the real-time tasks.(4) Resource co-allocation under reliability constraintAn extended reliability evaluation model is proposed, which is extended from existing Virtual Tree Grid Reliability Evaluation Model. In the extended model, a new type of failure is introduced to support reliability evaluation for real-time jobs in grid environments. By applying queueing system to model the working of grid resources, the proposed model is capable of supporting the resources with parallel processing capability, which is unsupported in previous model. The validity of the extended model and its approach to calculate real-time job's reliability are presented theoretically, and extensive simulations are conducted to verify its performance. Experimental results show that the proposed model can significantly improve the accuracy of reliability evaluation in presence of high-level workload. Also, it is able to reduce the jobs'mean response time due to its workload-aware feature.