Autoscheduling is a parallel program compilation and execution model that combines uniquely three features: Automatic extraction of loop and functional parallelism at any level of granularity, dynamic scheduling of pa...
详细信息
ISBN:
(纸本)3540649522
Autoscheduling is a parallel program compilation and execution model that combines uniquely three features: Automatic extraction of loop and functional parallelism at any level of granularity, dynamic scheduling of parallel tasks, and dynamic program adaptability on multi-programmed shared memory multiprocessors. This paper presents a technique that enhances the performance of autoscheduling in Distributed Shared Memory (DSM) multiprocessors, targetting mainly at medium and large scale systems, where poor data locality and excessive communication impose performance bottlenecks. Our technique partitions the application Hierarchical Task Graph and maps the derived partitions to clusters of processors in the DSM architecture. Autoscheduling is then applied separately for each partition to enhance data locality and reduce communication costs. Our experimental results show that partitioning achieves remarkable performance improvements compared to a standard autoscheduling environment and a commercial parallelizing compiler.
Application development for high-performance distributed computing systems, or computational grids as they are sometimes called, requires "grid-enabled" tools that hide mundane aspects of the heterogeneous g...
详细信息
NYNET (ATM wide area network testbed in New York state) Communication System (NCS) is a multithreaded message passing system developed at Syracase University that provides highperformance and flexible communication s...
详细信息
NYNET (ATM wide area network testbed in New York state) Communication System (NCS) is a multithreaded message passing system developed at Syracase University that provides highperformance and flexible communication services over asynchronous transfer mode (ATM) based highperformance distributed computing (HPDC) environments. NCS capitalizes on thread based programming model to overlap computations and communications, and develop a dynamic message passing environment with separate data and control paths. This leads to a flexible and adaptive message passing environment that can support multiple flow control, error control, and multicasting algorithms. We provide an overview of the NCS architecture and present how NCS point to point communication services are implemented. We also analyze the overhead incurred by using multithreading and compare the performance of NCS point to point communication primitives with those of other message passing systems such as p4, PVM, and MPI. Benchmarking results indicate that NCS shows comparable performance to other systems for small message sizes but outperforms other systems for large message sizes.
In order to solve real-time control problems, a good software design together with an appropriate control scheme and a system identification method are extremely importance. To facilitate the software design to cope w...
详细信息
In order to solve real-time control problems, a good software design together with an appropriate control scheme and a system identification method are extremely importance. To facilitate the software design to cope with such a time critical system, the concept of imprecise and approximate computation has been imposed and applied in real-time scheduling problems for more than a decade. Applying neural network to solve real-time problem is always a problem to neural network practitioners. In this paper, a principle for neural computation and real-time system 一 imprecise neural computation 一 will be presented. This principle extends the idea of imprecise computation in real-time systems by introducing concepts like mandatory neural structure and imprecise pruning. Using such concepts, it is able to design and analyze a real-time neural system for different real-time applications.
The proliferation of the high-speed network and powerful workstation has narrowed the performance gap between workstation clusters and *** draws significant attention of using Network of Workstations(NOW) environment ...
详细信息
The proliferation of the high-speed network and powerful workstation has narrowed the performance gap between workstation clusters and *** draws significant attention of using Network of Workstations(NOW) environment as parallel computing platform. Message Passing Interface(MPI) has been proposed as a standard for interprocess communication in message-passing parallel ***, it has language bindings for C,C++ and *** the Java technology appears to be able to unify the exist heterogeneous environment into one,we have implemented MPI in Java to exploit its usage in parallel computing *** this paper,we discuss the effectiveness of implementing MPI using *** also present some experimental results of our system called Java Message Passing Interface(JMPI).
In this paper,an extended Kalman filter(EKF) based training and data cleaning algorithm for neural network which can simultaneously train a neural network with good generalization and recover the original data from ...
详细信息
In this paper,an extended Kalman filter(EKF) based training and data cleaning algorithm for neural network which can simultaneously train a neural network with good generalization and recover the original data from noise will be *** with the standard recursive least square(RLS) training approach,it is demonstrated that the proposed algorithm can have a better generalization *** also present in this paper a parallel algorithm for the implementation of this EKF based training method onto a SIMD *** a n processors SIMD machine together with this parallel algorithm,the time required for one-step update can be reduced by an order of n,where n is the total number of hidden units,compared with using single sequential machine.
NYNET (ATM wide-area network testbed in New York state) Communication System (NCS) is a multithreaded message-passing tool developed at Syracuse University that provides low-latency and high-throughput communication s...
详细信息
NYNET (ATM wide-area network testbed in New York state) Communication System (NCS) is a multithreaded message-passing tool developed at Syracuse University that provides low-latency and high-throughput communication services over Asynchronous Transfer Mode (ATM)-based high-performance distributed computing (HPDC) environments. NCS provides flexible and scalable group communication services based on dynamic grouping and tree-based multicasting. The NCS architecture, which separates the data and control functions, allows group operations to be implemented efficiently by utilizing the control connections when transferring status information (e.g. topology information, routing information). Furthermore, NCS provides several different algorithms for group communication and allows programmers to select an appropriate algorithm at runtime. The authors overview the general architecture of NCS and present the multicasting services provided by NCS. They analyze and compare the performance of NCS with that of other message-passing tools such as p4, PVM, and MPI in terms of primitive performance and performance.
NASA's aggressive plans for extensive deep space exploration with a new generation of unmanned space craft requires a dramatic increase in responsibilities for onboard computing to reduce size, weight, and power t...
详细信息
Many sophisticated formalisms exist for specifying complex system behaviors, but methods for specifying performance and dependability variables have remained quite primitive. To cope with this problem, modelers often ...
详细信息
Many sophisticated formalisms exist for specifying complex system behaviors, but methods for specifying performance and dependability variables have remained quite primitive. To cope with this problem, modelers often must augment system models with extra state information and event types to support particular variables. This often leads to models that are non-intuitive, and must be changed to support different variables. To address this problem, we extend the array of performance measures that may be derived from a given system model by developing new performance measure specification and model construction techniques. Specifically, we introduce a class of path-based reward variables, and show how various performance measures may be specified using these variables. Path-based reward variables extend the previous work with reward structures to allow rewards to be accumulated based on sequences of states and transitions. To maintain the relevant history, we introduce the concept of a path automaton, whose state transitions are based on the system model state and transitions. Furthermore, we present a new procedure for constructing state spaces and the associated transition rate matrices that support path-based reward variables. Our new procedure takes advantage of the path automaton to allow a single system model to be used as the basis of multiple performance measures that would otherwise require separate models or a single more complicated model.
NASA's aggressive plans for extensive deep space exploration with a new generation of unmanned space craft requires a dramatic increase in responsibilities for onboard computing to reduce size, weight, and power t...
详细信息
NASA's aggressive plans for extensive deep space exploration with a new generation of unmanned space craft requires a dramatic increase in responsibilities for onboard computing to reduce size, weight, and power to drastically reduce mission costs. No longer will all raw sensor data be transmitted back to Earth but rather only the results from extensive preprocessing to decrease total link capacity by one to three orders of magnitude. A new approach to on board data processing is necessitated by this new expanded mission requirement. Processor-In-Memory (PIM) architecture combines logic and memory on the same integrated circuit to permit direct access to memory row buffers, greatly increasing effective memory bandwidth, reducing overhead and latency, and improving power efficiency. Organizing and managing the resources of PIM based systems is challenging. The PIM-MT architecture is an innovative approach to harnessing the physical resources of PIM technology through a message driven multithreaded virtual execution model. This paper describes the concepts of PIM-MT and its role in future spaceborne computing.
暂无评论