Summary form only given, as follows. this talk briefly reviews some of the most popular high-level and low-level parallel programming languages used for scientific computing. We will report our experiences of using th...
详细信息
ISBN:
(纸本)0769515738
Summary form only given, as follows. this talk briefly reviews some of the most popular high-level and low-level parallel programming languages used for scientific computing. We will report our experiences of using these languages in our research and compare the performance of several parallel scientific equation solvers implemented in different parallel languages. Major features and comparisons of these languages will be discussed. Some insights into when and where these languages should be used will be provided.
Current data management solutions are optimized for intra-enterprise, client-server applications. they depend on predictability, predefined structure, and universal administrative control, and cannot easily cope with ...
详细信息
ISBN:
(纸本)0769514804
Current data management solutions are optimized for intra-enterprise, client-server applications. they depend on predictability, predefined structure, and universal administrative control, and cannot easily cope with change and lack of structure. However, modern e-commerce applications are dynamic, unpredictable, organic, and decentralized, and require adaptability. eXtensible Data Management (XDM) is a new approach that enables rapid development and deployment of networked, data-intensive services by providing semantically-rich, high-performance middle-tier data management, and allows heterogeneous data front different sources to be accessed in a uniform manner. Here, we discuss how middle tier extensible data management can benefit an enterprise, and present technical details and examples from the Index Fabric, an XDM engine we have implemented.
Recent increases in CPU performance have outpaced increases in hard drive performance. As a result, disk operations have become more expensive in terms of CPU cycles spent waiting for disk operations to complete. File...
详细信息
Recent increases in CPU performance have outpaced increases in hard drive performance. As a result, disk operations have become more expensive in terms of CPU cycles spent waiting for disk operations to complete. File prediction can mitigate this problem by prefetching files into cache before they are accessed However, incorrect prediction is to a certain degree both unavoidable and costly. We present the Program-based and User-based Last n Successors (PULnS) file prediction model that identifies relationships between files through the names of the programs and the users accessing them. Our simulation results show that, in the worst case, PULnS makes at least 20% fewer incorrect predictions and roughly the same number of correct predictions as the last-successor model.
Peer-to-peer storage architecture aims to aggregate the storage of individual computers (peers) to form a universal shared storage space in which all peers undertake identical roles. the fully distributed nature of su...
详细信息
ISBN:
(纸本)0769515738
Peer-to-peer storage architecture aims to aggregate the storage of individual computers (peers) to form a universal shared storage space in which all peers undertake identical roles. the fully distributed nature of such an architecture makes it possible to support features such as scalability at a global scale, self-configuration and dynamic adaptation, permanence, fault tolerance and anonymity. Representative examples include Freenet, OceanStore, PAST (Peer-to-peer Archival Storage Tool) and CFS (Cooperative File System). In this paper, we study the memory architecture of such systems, especially the effects of caches and directories on their performance. To facilitate the study, an abstract model, called the distributed shared memory (DSM) model, is first proposed to capture the essence of the peer-to-peer storage architecture from the memory perspective. Six variations representing different points in the peer-to-peer storage design space are then identified. three state-of-the-art peer-to-peer storage systems are cast on to these models and then examined qualitatively. the performance of these models under different memory pressures, network sizes and failure degrees is finally evaluated via simulation.
We propose a hybrid parallelism-independent scheduling method, predominantly performed at compile time, which generates a machine code efficiently executable on any number of workstations or PCs in a cluster computing...
详细信息
We propose a hybrid parallelism-independent scheduling method, predominantly performed at compile time, which generates a machine code efficiently executable on any number of workstations or PCs in a cluster computing environment. Our scheduling algorithm called the dynamical level parallelism-independent scheduling algorithm (DLPIS) is applicable for distributed computer systems because additionally to the task scheduling, we perform message communication scheduling. It provides an explicit task synchronization mechanism guiding the task allocation and data dependency solution at run time at reduced overhead. Furthermore, we provide a mechanism allowing the self-adaptation of the machine code to the degree of parallelism of the system at run-time. therefore our scheduling method supports the variable number of processors in the users' computing systems and the adaptive parallelism, which may occur in distributed computing systems due to computer or link failure.
A small number of network components failures can cause high-speed local area networks (LANs) or System Area Networks (SANs) to be partitioned. this issue has been resolved withthe introduction of redundant network c...
详细信息
A common trend in the design of large-scale clusters is to use a high-performance data network to integrate the processing nodes in a single parallel computer. In these systems the performance of the interconnect can ...
详细信息
Checkpointing is a key technology for applications on large cluster computer systems. As cluster sizes grow, component failures will become a normal part of operation, and applications will have to deal more directly ...
详细信息
Checkpointing is a key technology for applications on large cluster computer systems. As cluster sizes grow, component failures will become a normal part of operation, and applications will have to deal more directly with repeated failures during program runs. We describe automatic checkpointing in the ZPL compiler and its advantages over traditional library or system-based approaches that have no information about application behavior. We show that even naive compiler-inserted checkpoints can significantly reduce the size of the checkpoint recovery data, up to 73% in our application suite. We also introduce the notion of checkpoint ranges, a range of code where processors can perform a local checkpoint at any time during the range. the compiler guarantees that these local checkpoints form a globally consistent checkpoint without global coordination by ensuring that there are no in-flight messages during the checkpoint range. Checkpoint ranges help further alleviate any additional network congestion caused by checkpointing.
暂无评论