Dingxing Wang

Distributed page ranking in structured P2P networks

We discuss the techniques of performing distributed page ranking on top of structured peer-to-peer networks. Distributed page ranking are needed because the size of the Web grows at a remarkable speed and centralized page ranking is not scalable. Open system PageRank is presented based on the traditional PageRank used by Google. We then propose some distributed page ranking algorithms, partially prove their convergence, and discuss some interesting properties of them. Indirect transmission is introduced to reduce communication overhead between page rankers and to achieve scalable communication. The relationship between convergence time and bandwidth consumed is also discussed. Finally, we verify some of the discussions by experiments based on real datasets

Improving Availability of P2P Storage Systems

This paper discusses how to building high available storage systems on top of peer-to-peer infrastructure. We first demonstrate that traditional definitions of availability are not suitable for distributed systems. Then the application-specific availability is defined and analyzed in this paper. We observed by experiments that, in a P2P storage system with frequently node join and leave, availability can’t be improved effectively by pure redundancy (replication or erasure coding). A method is proposed in this paper which combines peer auto-selection and redundancy to effectively improve the availability of P2P storage systems.

DSI: distributed service integration for service grid

This paper presents DSI, a distributed service discovery and integration utility for service grid. The goal of DSI is efficiently to improve service publishing, deletion and lookup, decentralized control, scalability, and availability. DSI comprises a logically global metadata pool and a double-layered DHT (Distributed Hash Table). Efficient service publishing and lookup are achieved by adopting DHT. Clean service deletion is guaranteed by time-stamping. And system availability is improved by replication combined with automatic indexer selection. Experiments show that DSI can achieve short response time, low processing cost and high availability.

BUSTER: an integrated debugger for PVM

Parallel debugging focuses on solving four problems: nondeterminism, global states, communication related errors and performance debugging. PVM is a widely used programming environment for network of workstations, but it provides little support on debugging. The authors have developed an integrated parallel debugger (called BUSTER) for PVM programs running on workstation clusters. They introduce the ideas and implementation techniques of the debugger.

Study on topologies of Information Grid

Information Grid is the combination of information integration and the Grid technology. It makes higher level of data abstraction than traditional computation/data grid. This paper defines a logic component model for information grid. Based on this, three possible topologies for information grid are proposed and studied. Four measures (scalability, availability, maintainability, and information retrieval performance) are used to compare and analyze these three topologies.

CHECKPOINTING AND ROLLBACK RECOVERY FOR NETWORK OF WORKSTATIONS

Network of workstations (NOW) now becomes one of the main trends of parallel computing. But for long-running scientific programs, it needs effective fault tolerance for its changing property. Checkpointing and rollback recovery is a solution to this problem. First the main problems upon rollback recovery are discussed, the different checkpointing techniques for NOW are analyzed, and then the design and implementation of ChaRM (checkpoint-based rollback recovery and process migration) system are described. The comparison of three coordinated checkpointing systems is given.

Granularity analysis for exploiting adaptive parallelism of declarative programs on multiprocessors

Declarative Programming Languages (DPLs) apply a process model of Horn clauses such as PARLOG[8] or a reduction model of λ-calculus such as SML[7] and are, in principle, well suited to multiprocessor implementation. However, the performance of a parallel declarative program can be impaired by a mismatch between the parallelism available in an application and the parallelism available in the architecture. A particularly attractive solution is to automatically match the parallelism of the program to the parallelism of the target hardware as a compilation step. In this paper, we present an optimizing compilation technique called granularity analysis which identifies and removes excess parallelism that would degrade performance. The main steps are: an analysis of the flow of data to form an attributed call graph between function (or predicate) arguments; and an asymptotic estimation of granularity of a function (or predicate) to generate approximate grain size. Compiled procedure calls can be annotated with grain size and a task scheduler can make scheduling decisions with the classification scheme of grains to control parallelism at run-time. The resulting granularity analysis scheme is suitable for exploiting adaptive parallelism of declarative programming languages on multiprocessors.

Modeling replication strategies in data grid systems with arbitrary clustered demands

This paper considers the relationship between request distribution and replica distribution in data grid when request exhibits arbitrary clustered demands. We first give formal model of replication strategies in data grid system. Second, we investigate what is optimal way at the objective of minimizing average access latency to replicate data when request exhibits arbitrary clustered demands. We explain why replicas should be replicated uniformly when request is uniformly distributed in a sub grid in the sense of optimal strategy. Then we investigate the relationship between different files in a sub grid. Furthermore, we analyze the case when all sub grids are equal-sized and conclude that when request is uniformly distributed in system, replicas should be uniformly distributed in system too. Finally, we give an optimal strategy when sub grids are not equal-sized and different sub grids exhibit different request clustering patterns. Compared with some popular strategies, the optimal strategy has some advantages of lower wide area network bandwidth requirement and lower average access latency. Simulation results validate the effectiveness of optimal strategy.

Impact of Clustered Demands on Performance of Replication Strategies in Data Grid Systems

This paper considers the impact of clustering demands on download performance of data grids. The performance metrics are hit ratios and average access latency. For a replication strategy, we build a mechanism by means of majorization theory to compare system performance for two clustering patterns. We employ proportional replication strategy as an example to illustrate the effectiveness of this mechanism. We find that clustering can increase local site hit ratio of proportional replication strategy. However, clustering does not always bring benefits to system. The inter grid hit ratio and average access latency will deteriorate for some cases. If the function of inter grid hit ratio is convex on the domain of file request rate, clustered demands will increase inter grid hit ratio compared with uniform demand. Simulation results validate the correctness of the trend deduced by this mechanism.

Hierarchical Replica Location Service Based on Hybrid Overlay Platform

An important problem that confronts data grid applications is to efficiently locate the node that stores a particular data replica. This paper proposes a hierarchical replica location service (HRLS) based on a hybrid overlay platform to address this problem. The system architecture of the HRLS is defined recursively. The granularity of nodes can be tuned flexibly according to system needs. Lookup messages can be resolved level by level moving gradually towards the destination node. There are two kinds of metadata; one is user-defined and the other is system-oriented. Every kind of data has its own bidirectional mapping rule for the metadata. This also helps system security because mapping rules and data can only be accessed by authenticated users. The HRLS has been implemented in an enterprise data grid which was spread over four cities. Results from measurement and the implemented system show that the HRLS successfully met the design goals