Research On Adaptive Resource Management For Service Systems

With the matures of service computing and the emergence of new service paradigms, such as mobile internet, Internet of Things and Software as a Service, service systems have evolved to multi-tenant complicated system with high concurrent and multiple device access. Meanwhile, most service systems have been deployed on exclusive infrastructure and the expansion of service systems rely on the unceasing infrastructure investment, which decreases resource utilization and increases the operation cost of service systems. Service systems require more automatic mechanism to consolidate various resources and more elastic strategy to schedule resource, in order to increase total resource utilization, meet the QoS requirement of various service components and improve the efficiency of service deliveryCloud Computing is the significant transformation of IT service provisioning paradigm. In this paradigm, service and data are stored in scalable shared resource pool via virtualization technology, and can be elastically provisioned to adapt to users’ demand, which introduced novel delivery solution for the design, development and deploy of service systems. Based on virtualization technology and the concept of autonomic computing, this paper studies adaptive resource management for service systems. We propose the architecture of adaptive resource management and study four related issues, that is, resource discovery, resource deployment, resource provision and resource negotiation. We construct self-organization collaborated resource overlay, design efficient virtual machine placement mechanism and dynamic resource provisioning strategy, and build up adaptive resource negotiation and bidding model, to achieve automatic virtual resource management. The proposed resource management architecture would guarantee the quality of service delivery and improve total resource utilization.The major contribution of this thesis is as following:1) According to bottleneck of performance and point of failure introduced by centralized management of resource pool, we propose a Gossip-based Hybrid Multi-attribute Overlay (GHMO) to construct self-organization resource infrastructure with peer-to-peer fashion. GHMO maintain both of structured and unstructured topology, and use hybrid routing algorithm to implement multi-attribute accurate query and range query, which would raise the efficiency of resource query. Meanwhile, a weight overlay is introduced in the case of network latency, and a nearest neighbor selection strategy is raised to optimize the choice of query routing by selecting neighbor nearby to transmit query. Simulation results show that, compared with Multi-Attribute Hybrid Overlay (MAHO), query efficiency of GHMO is increased by33%. Also, the routing cost is reduced by14.2%with nearest neighbor selection strategy.2) To deal with server sprawl and high energy consumption caused by unbalanced allocation of multiple resources in resource pool, we propose Power-aware Heuristic Vector Placement (PHVP) mechanism in heterogeneous cloud scenarios. The proposed mechanism considers multiple dimensions of the capacity of physical machines and the demand of virtual machines (VM), and utilizes multiple dimension collaboration to place virtual machines, in order to achieve server consolidation and reduce energy consumption. PHVP apply Greedy Randomized Adaptive Search Procedures (GRASP) to reduce of local optima introduced by classical heuristic methods and improve the efficiency of VM placement. Also, power-aware best-fit host activation algorithm is introduced to optimize the usage of physical machines in heterogeneous resource pool. Simulation results show that, in heterogeneous resource pool, PHVP outperforms FFD and vector based approach, decreasing amount of physical machines and total power consumption by10%and10.3%respectively.3) To satisfy the demand of high concurrent and deadline constraint computing tasks, we propose a dynamic resource provisioning strategy for deadline constraint tasks. The proposed strategy adjusts the amount of virtual machines in resource pool and the order of task execution dynamically, to adapt to varying request workload and fulfill different level of QoS. We design analytical provision model Elastic Resource Provisioning Model (ERPM) for adaptive provision based on queuing theory, by adjusting the amount of virtual machines in resource pool periodically based on the status of task queue. Also, we schedule the order of task execution based on the weight of each priority calculated by Weighted Fair Queuing (WFQ) algorithm and SLA violation cost. Simulation results show that, ERPM give elastic resource provisioning for dynamic workload. Also, compared with FCFS, SJF and EDF, WFQBS algorithm could fairly schedule tasks with different priority, and reduce the total cost of SLA violation.4) To cope with various demand/supply amounts and different trade targets of resource provider and resource consumer in service resource market, we propose resource bidding model based on Bundle Multiunit Double Auction (BMDA). The bidding model would determine resource allocation scheme by bidding among buyer and seller, to inspire trade enthusiasm of market participants and maximum social utility. Based on this model, we propose two auction mechanism, linear programming based mechanism BMDA-LP and greedy algorithm based mechanism BMDA-GREEDY. These two auction mechanisms are proved to be strategy-proof, budget-balanced and individual rational. Simulation results show that, two auction mechanisms could regulate auction market adapted to resource demand/supply. Moreover, the auction efficiency of BMDA-LP outperforms that of BMDA-GREEDY, with the growth of161%,70%and37%in social utility, user satisfaction degree and resource utilization respectively...