Researches On The Key Technologies Of High Performance Computing Based On Structured Peer-to-Peer Network

Researches about High Performance Computing based on structured peer to peer network are now still in its initial stage. A bunch of new problems are arising, such as the optimization of the topology, range queries in DHT based network, task scheduling in the platform and etc. Solutions to these problems are bases of a practicable platform. Churn is a basic characteristic of P2P network. It means the nodes join and leave the system at an uncertain time. That makes the system difficult to design. In order to achieve the whole platform, some key topics will be discussed.For a DHT based peer-to-peer network environment, the efficiency of keyword queries is an important mark. The Chord protocol builds a structured hash table on a ring space. When queries began, it follows a clockwise direction. By extending the router table of Chord, the new designed Dual-Chord protocol proposed a bidirectional search strategy. With this strategy, queries can go in an anticlockwise direction as needed, and the average query path will be much shorter. On the other hand, in Dual-Chord, network latency is designed to be a parameter of the algorithm, which helps to select an optimal path. Experiment results showed that Dual-Chord is more effective than Chord in searching process.Range query module is an infrastructure of the whole HPC platform which based on structured peer-to-peer network. To achieve range queries in DHT, Yarqs, a new scheme is proposed. Yarqs catch the data route the node and organize the data into the local cache. The caches on every node then form a network. Range queries are deployed on the cache network. Such schema of range queues has excellent adaptability to Churn P2P environment. Yarqs can get high accuracy and also support inherent load balance. Many experiments are designed to examine the performance of Yarqs. And it is proved that to support effective range queries, Yarqs is a good protocol.Based on the optimized overlay and the infrastructure of range queries module, the architecture of resource management and resource discovery of the whole system is built. In the architecture, the monitor module of the system is first discussed, and then the different logic levels of the query sub-system. In each level of the module, the function and interface of each level are clearly defined. On the other hand, to verify the efficiency of this module, the attributes of nodes in a real trace of a P2P network are analyzed and some statistics characters of the nodes have been gotten. At the same time, Yarqs is extended to support multi-attribute range queries.P2P environment is a heterogeneous and asynchronous environment. At the same time, P2P platform is not stable. The joining and leaving of peers are random. These characteristics make the P2P based HPC platform have great difference to the traditional HPC platform and the global computing projects. To achieve effective task scheduling on P2P based platform, a DHT based task management scheme is introduced in. Further, a data structure of distributed bidirectional Skiplist is developed to keep the prediction session time. The new task scheduling strategy tries to distribute the task to the nodes which have longer online session time. With such scheme, the scheduling strategy can reduce the migration of tasks among different nodes and improve the resource utilization of computing nodes.