Showing papers on "SPMD published in 1988"

Image processing on a SIMD/SPMD architecture: OPSILA

Abstract: OPSILA, is a description given of a general-purpose parallel architecture with two different forms of parallelism: the well-known SIMD (single-instruction, multiple data shown), a synchronous form of parallelism; and SPMD (single program, multiple data stream), which is an asynchronous mode. It is shown that OPSILA is efficient for a wide variety of image algorithms including low and high level processing. The efficiency of OPSILA is demonstrated for the low-level algorithms, through the implementation of a set of typical operations: local (convolution), global (histogram), and geometric corrections. >

On the Effects of Synchronization in Parallel Computing

Abstract: We present a non-deterministic model of parallel computation that includes the effects of communication costs, computation control costs and synchronization effects. Synchronization may be the most important effect in many important applications. Our model is particularly suited for coarse grain parallelism, as in Same Program Multiple Data (SPMD) computations. Using this model we derive exact expressions for synchronization costs, where the parallel tasks have execution times that are unifonnly or exponentially distributed. We show that efficient massive parallelism is possible with the unifonn distribution, but the synchronization costs for exponentially distributed execution times lead to a logarithmically decayiog efficiency.

Finite Element Analysis on a Highly Parallel Multiprocessor Architecture

Abstract: We consider the use of new parallel computer architectures for the computationally intensive parts of finite element algorithms. The first part, the assembly of the finite element equations, is a purely local procedure and is ideally suited to parallel computation. The second part, the solution of the nodal equations, involves the simultaneous solution over the whole domain. The problem is partitioned between the processors using a domain decomposition which specified the matrix structure to be used in the solution. We assume that the processors will be operating in SIMD (single instruction multiple data) mode in which they all obey the same sequence of instruction but operate on separate data sets or in SPMD (single program multiple data) mode in which they all loaded will the same program but may take different logical paths through the code. We consider a number of preconditioned conjugate gradient schemes with optimal and non-optimal preconditioners.

Non-Algorithmic Load Imbalance Effects for Domain Decomposition Methods on a Hypercube

Abstract: In this paper we discuss the effects of non-algorithmic load imbalance for three synchronization structures which appear in different algorithms using domain decomposition methods for solving PDE's on a hypercube machine. The synchronization structures are determined from the conditions related to the partitioning of a hypercube. To characterize the load imbalance effects we introduce a factor 'V defined as the ratio of the load imbalance costs to the total parallel execution time in absence of any load imbalance. We develop non-deterministic models for SPMD (Same Program Multiple Data) execution model and compute the 'V factor for the three structures, for different types of distributions of the execution time.