摘要:

This paper examines the structure of artificial neural networks (ANN) and the operation of their algorithms in order to identify the forms of parallelism that may be inherent in them. Parallelism within the topological structure of ANNs are seen to be of two forms: neuron and synapse. Operational parallelism is also of two forms: training set parallelism and recall/teaching parallelism. Performance models are formulated to predict the likely speed improvement achieved due to parallelism.

ISSN:1063-7192    

DOI:10.1080/10637199708915593

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

In this paper parallel algorithms for fractal image compression on MIMD architectures are introduced, classified, and discussed. The crucial point for the choice of a suitable parallelization strategy is the memory capacity of a processor element of the target architecture, therefore the classification follows the memory demand of the algorithms. We present experimental results of all main-representatives of the different algorithm classes. The results show a decreasing speedup for increasing memory demand for most of the proposed algorithms.

ISSN:1063-7192    

DOI:10.1080/10637199708915594

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

This paper deals with the parallel computation of two-dimensional Navier-Stokes equations in the vorticity stream function formulation. The classical Schwarz alternating method is used to solve the evolution convection-diffusion equation and an efficient variant of this method to solve the associated Poisson equation. Both previous methods are implemented on a distributed memory multiprocessor. The test case of the driven square cavity is chosen for discussing the efficiency of these coupled methods.

ISSN:1063-7192    

DOI:10.1080/10637199708915595

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

In this paper, a linear systolic array is developed for the conjugate gradient (CG) and preconditioned conjugate gradient (PCG) methods. By reformulating the algorithms into single assignment code, the resulting dependence graphs (DG) are regularized and then mapped onto a signal flow graph (SFG) and systolic array developed. The obtained linear systolic array can achieve a speedup ofO(n) when n processing elements are used.

ISSN:1063-7192    

DOI:10.1080/10637199708915596

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

In this paper we study the implementation of a variant of the classic Gauss-Jordan (GJ) method which was recently introduced by Huard [8] on a shared memoryMIMDcomputer. Two parallel versions are derived by dividing the sequential Huard method into noninterfering tasks. Taking into consideration the computation as well as the communication complexity we present a parallel scheduling algorithm for each task graph. Next, in an attempt to reduce the communication cost we introduce block versions and follow a similar approach for their study.

ISSN:1063-7192    

DOI:10.1080/10637199708915597

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

This paper represents a new efficient parallel implementation of neural networks on mesh-connected SIMD machines. A new algorithm to implement the recall and training phases of the multilayer feedforward network with backpropagation is devised. The developed algorithm is considered much faster than other known algorithms; it requires O(l) multiplications and O(logN) additions, whereas most others require O(N) multiplications and O(N) additions. In this paper we restrict the algorithm to map a neural network of a maximum of n neurons per layer on anN × Nmesh of processors whereN ≥ n, however, it may be extended to the general case. Time comparisons with other algorithms are furnished.

ISSN:1063-7192    

DOI:10.1080/10637199708915598

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

Multicast is an important interprocessor communication pattern existing in various parallel application algorithms. This paper proposes a new multicast algorithm for hypercube computers that not only minimizes the number of hops required for delivery of the message and keeps the amount of traffic small, but also takes into consideration of balancing traffic among the active processors to maximize the parallelism in communication. Our algorithm has a time complexity ofO(nN) for fault-tolerant multicasting in a hypercube ofN = 2nprocessors containing an arbitrary number of faulty processors.

ISSN:1063-7192    

DOI:10.1080/10637199708915599

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

The Lanczos algorithm is one of the most widely used methods for finding a small number of the extremal eigenvalues and associated eigenvectors of large, sparse, symmetric matrices. In this paper the performance on two parallel machines with different architectures of a modified version of the algorithm which incorporates a novel convergence monitoring method is assessed. The investigation has been carried out using a shared memory Convex C3840 with two processors and a 16-node Intel iPSC/860 hypercube. It is shown that parallel implementations of the modified algorithm can efficiently exploit the facilities provided by both machines. However, there are significant architecture dependent considerations which favour the use of the shared memory machine for the solution of general instances of the problem. These considerations relate to the cost of inter-processor communication and the limited availability of fast memory on the distributed memory machine.

ISSN:1063-7192    

DOI:10.1080/10637199708915600

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor

摘要:

Two different strategies are discussed, for the partitioning of composite grids, in the context of parallel computers of MIMD type with distributed memory. The two strategies are compared theoretically, with respect to arithmetic load balance and communication overhead. Moreover, the ability to handle dynamically changing grids is considered.

ISSN:1063-7192    

DOI:10.1080/10637199708915601

出版商:Taylor & Francis Group    

年代:1997

数据来源: Taylor