Research On Resource Discovery And Allocation In Semi-distributed Systems

With the rapid development of computer and network technology, the demand ofresource and service has been urgently raised. How to organize resource distributedaround the world and provide service becomes the urgent requirement. As the hugecomputing scale， the complex network structure and the diverse demand, efficientresource discovery, reasonable job allocation and resource collaboration have been keyproblems of distributed computing.The topology of distributed system includes centralized system, fully distributedsystem and semi-distributed system. The resource management of semi-distributedsystem is discussed in this paper. Resource management is the connection of resourceprovider and user. It includes resource discovery, resource monitor, resource allocation,resource migration and fault tolerant. Resource discovery is the first and the mostimportant problem in resource management. It aims at returning the identifier ofresource set according the request of user. It helps users get the needed informationfrom huge information. The key problem of resource discovery is locating resourcerapidly and efficiently with low redundant message and system overhead. Resourceallocation focuses on improving the efficiency of resource and increasing systemthroughput. It also concerns on reducing waiting time and improving load balance.This dissertation focuses on resource discovery and allocation in semi-distributedsystems. The major contributions are as follows:1. A group division method based on the relationships of nodes in semi-distributedsystem was presented. The relationship includes the communication distance,collaboration and trust degree. This method provides an elegant aggregate solution sothat nodes with best mutual trust and closest co-operation have more probability to bedivided into the same logic group. The relationship is described by PRG(peerrelationship graph) and the group divison tool is Min-max cut. The major advantage isthat our methodology can reduce group cross query, which optimizes the response timeand recall rate.2. A resource allocation method based on sentinel agent and queuing theory was proposed. Sentinel agent carries on user request and wanders in the network. The pathof sentinel agent is decided by the decision model. We use M/M/C/queuing model toimprove the resource allocation. The major advantage is that it reduces the migrationtimes and improves the successful rate. Also it is dynamic and scalable.3. A resource allocation method based on Linear Programming model waspresented to improve the priority of user and request. It uses job clustering to coordinateresource allocation. The optimal allocation strategy was gained by the best value ofdifferent target functions. Through the dynamic monitoring for system load andbandwith, the representative improves the accuracy for the task execution time andcommnication time estimation, which increases the degree of parallelism.4. A novel node feedback and incentive mechanism based on local trust degree andglobal trust degree was introduced. Local trust degree is described by the feedback oftwo nodes with collaboration. Global trust degree is decided by a node’s historybehavior, which includes response time, waiting time and successful rate. The resourceallocation and selection focus on both the two factors. According to the activity andcontribution degree of nodes, a node incentive mechanism based on multi-stagedecision-making was presented to solve the selfish user problem.5. A multi-layer semi-distributed system is implemented for resource discovery andallocation. It contains the advantage of self-adaptive and self-organizing model. Eachlayer is independent with good Scalability. The system provides dynamic node regroupand decision making mechanism, which improves the efficiency of resource query andload balance.