Instruction Multiple Data Research Articles

A Simulation Tool for the Performance Evaluation of a Class of Mimd Systems

A user-friendly, menu-driven simulation tool for the performance evaluation of a class of Multiple Instruction Multiple Data (MIMD) systems with private caches is presented in this paper. This tool can be efficiently utilized in the preliminary stages of designing such systems. The class of MIMD machines modelled consists of processors, private caches, and a set of interleaved memory modules. Each processor has a local private cache which is connected through a cross-bar interconnection network to shared Main Memory modules. Specifications of elements can be denned individually or globally to accommodate various homogeneous/heterogeneous system configurations. The tool can be used to configure, model, and gimnljfcto the MIMD computer system and its performance. It uses a set of metrics to indicate the model's performance by estimating specified performance indices of the system and its components that include relative speedup, throughput, and utilization as well as others such as average queue length and average waiting time. Simulation runs were performed to validate and assess the tool's functionality.

Aerodynamic Influence Coefficient Computations Using Euler/Navier-Stokes Equations on Parallel Computers

An efficient procedure to compute aerodynamic influence coefficients (AIC) using high fidelity flow equations such as Euler/Navier- Stokes equations is presented. The AIC's are computed by perturbing structures using mode shapes. The procedure is developed on a multiple-instruction, multiple-data (MIMD) parallel computer. In addition to discipline parallelization and coarse-grain parallelization of the flow domain, embarrassingly parallel implementation of ENS AERO code demonstrates linear speedup for a large number of processors. Demonstration of the AIC computation for static aeroelasticity analysis is made on an arrow wing-body configuration. The computations show that some of the AIC's do not converge at the lower perturbation range since the perturbation is too small to prevail over the initial aerodynamic loads. The effect of initial aerodynamic loads disappears with increasing perturbation amplitude. The demonstrated linear scalability for multiple concurrent analyses shows that the threelevel parallelism in the code is well suited for the computation of the AIC's.

On the Optimization of a Pipeline Model to Integrate a Reduced-Dimensionality Schrödinger Equation for Distributed Memory Architectures

This paper examines the problem of dealing with the integration of a two-dimensional Schrödinger equation using a sector-diabatic coupled-channel approach when the coupling matrix is too large to fit into the memory of individual nodes of a multiple-instruction, multiple-data (MIMD) parallel machine. For this case, the distribution of blocks of the coupling matrix among the different nodes and the use of a pipeline model are investigated. To measure performances and to compare them with those measured for a previously optimized task farm model, the pipeline model was implemented on a hypercube and its structure optimized. Benchmarks were performed for dimensions of the coupling matrix manageable by both task farm and pipeline models.

Distributed‐ and shared‐memory parallelizationsof assignment‐based data association formultitarget tracking

To date, there has been a lack of efficient and practical distributed‐ and shared‐memoryparallelizations of the data association problem for multitarget tracking. Filling this gap is oneof the primary focuses of the present work. We begin by describing our data association algorithmin terms of an Interacting Multiple Model (IMM) state estimator embedded into anoptimization framework, namely, a two‐dimensional (2D) assignment problem (i.e., weightedbipartite matching). Contrary to conventional wisdom, we show that the data association (oroptimization) problem is not the major computational bottleneck; instead, the interface to theoptimization problem, namely, computing the rather numerous gating tests and IMM stateestimates, covariance calculations, and likelihood function evaluations (used as cost coefficientsin the 2D assignment problem), is the primary source of the workload. Hence, for both ageneral-purpose shared‐memory MIMD (Multiple Instruction Multiple Data) multiprocessorsystem and a distributed‐memory Intel Paragon high‐performance computer, we developedparallelizations of the data association problem that focus on the interface problem. For theformer, a coarse‐grained dynamic parallelization was developed that realizes excellent performance(i.e., superlinear speedups) independent of numerous factors influencing problemsize (e.g., many models in the IMM, denseycluttered environments, contentious target‐measurementdata, etc.). For the latter, an SPMD (Single Program Multiple Data) parallelization wasdeveloped that realizes near‐linear speedups using relatively simple dynamic task allocationalgorithms. Using a real measurement database based on two FAA air traffic control radars, weshow that the parallelizations developed in this work offer great promise in practice.

MIMD vs. SIMD parallel processing: A case study in 3D medical image registration

This paper compares and contrasts the issues involved with implementing computationally intensive medical imaging algorithms on a multiple-instruction, multiple-data (MIMD) and single-instruction, multiple-data (SIMD) parallel processing computers. Implementation issues and timing analysis are presented for a 3D medical image registration algorithm implemented on a 16 processor SGI Challenge MIMD computer and on a 128 × 128 processor MasPar SIMD computer. The MIMD implementation is shown to have nearly N-times performance improvement with respect to a single processor when using N processors. It is also shown that the MIMD implementation is a minimum of four times faster than the SIMD implementation.

A parallel contact detection algorithm for transient solid dynamics simulations using PRONTO3D

An efficient, scalable, parallel algorithm for treating material surface contacts in solid mechanics finite element programs has been implemented in a modular way for multiple-instruction, multiple-data (MIMD) parallel computers. The serial contact detection algorithm that was developed previously for the transient dynamics finite element code PRONTO3D has been extended for use in parallel computation by utilizing a dynamic (adaptive) load balancing algorithm. This approach is scalable to thousands of computational nodes1

Parallel computing for chromosome reconstruction via ordering of DNA sequences

In this paper we present our practical experience with the design and implementation of a suite of parallel algorithms called PARODS, for chromosome reconstruction via ordering of DNA sequences. PARODS contains parallel algorithms based on an earlier serial algorithm ODS which is a physical mapping algorithm based on simulated annealing. The parallel simulated annealing (PSA) algorithms for physical mapping in PARODS are based on Markov chain decomposition. We propose and describe perturbation methods and problem-specific annealing heuristics in the context of the physical mapping problem. We describe implementations of parallel Single Instruction Multiple Data (SIMD) algorithms on a 2048-processor MasPar MP-2 system and implementations of parallel Multiple Instruction Multiple Data (MIMD) algorithms on an 8-processor Intel iPSC/860 system and on a cluster of UNIX workstations using the Parallel Virtual Machine (PVM) system. We discuss and analyze the convergence, speedup and scalability characteristics of the aforementioned algorithms. We show the best SIMD algorithm to have a speedup of 1232 on the 2048 processor MasPar MP-2 system and the best MIMD algorithm to have a speedup of 5.35 on the 8-processor Intel iPSC/860 system and 3.40 on a 6-workstation cluster running PVM.

A MIMD-based video signal processing architecture suitable for large area integration and a 16.6-cm/sup 2/ monolithic implementation

The architecture and implementation of a programmable video signal processor dedicated as building block of a multiple instruction multiple data (MIMD)-based bus-connected multiprocessor system is presented. This system can either be constructed from several single processor chips, or it can be integrated on a large area integrated circuit containing several processors. The processor allows an efficient implementation of different video coding standards like H.261, H.263, MPEG-1 and MPEG-2. It consists of a RISC processor supplemented by a coprocessor for computation intensive convolution-like tasks, which provides a peak performance of more than 1 giga-arithmetic operations per second (GOPS). A large area integrated circuit integrating 9 processor elements (PE's) on an area of 16.6 cm/sup 2/ has been designed. Due to yield considerations redundancy concepts have been implemented, that-even in the presence of production defects-result in working chips utilizing a lower number of PE's. Each PE has built-in self-test (BIST) capabilities, which allow for an independent test of itself under the control of its integrated fault-tolerant BIST controller. Defective PE's are switched off. Only the PE's passing the BIST are used for video processing tasks. Prototypes have been fabricated in a 0.8 /spl mu/m complementary metal-oxide-semiconductor (CMOS) process structured by masks using wafer stepping with overlapping exposures. Employing redundancy, up to 6 PE's per chip were functional at 66 MHz, thus providing a peak arithmetic performance of up to 6 GOPS.

Development of a parallel molecular dynamics code on SIMD computers: Algorithm for use of pair list criterion

In recent years several implementations of molecular dynamics (MD) codes have been reported on multiple instruction multiple data (MIMD) machines. However, very few implementations of MD codes on single instruction multiple data (SIMD) machines have been reported. The difficulty in using pair lists of nonbonded interactions is the major problem with MD codes for SIMD machines, such that, generally, the full connectivity computation has been used. We present an algorithm, the global cut-off algorithm (GCA), which permits the use of pair lists on SIMD machines. GCA is based on a probabilistic approach and requires the cut-off condition to be simultaneously verified on all nodes of the machine. The MD code used was taken from the GROMOS package; only the routines involved in the pair lists and in the computation of nonbonded interactions were rewritten for a parallel architecture. The remaining calculations were performed on the host computer. The algorithm has been tested on Quadrics computers for configurations of 32, 128, and 512 processors and for systems of 4000, 8000, 15,000, and 30,000 particles. Quadrics was developed by Istituto Nazionale di Fisica Nucleare (INFN) and marketed by Alenia Spazio. © 1998 John Wiley & Sons, Inc. J Comput Chem 19: 685–694, 1998

On parallelizing a groundwater pollution simulator

Domain decomposition strategies and computational mesh reordering are discussed for finite difference parallel simulations of groundwater contaminants transport. The parallel performance of point iterative methods traditionally used in groundwater pollution modelling is studied. The algorithms were implemented with red-black and wavefront reordering of the computational mesh. A standard conservative transport equation defined on a two-dimensional grid with Dirichlet boundary conditions was used for the analysis. Completely portable multiple instructions multiple data (MIMD) implementations of the algorithm were performed using message-passing interface (MPI). The runtimes of the algorithms are presented as a function of grid refinement and number of processors, and the communication overhead of the parallel simulation process is investigated, showing that the red-black reordering technique yields the best performance results. The method also provides higher efficiency and scalability when applied to large-scale problems. Optimal parameters are suggested for parallel simulation of groundwater pollution using finite difference schemes.

World Wide Web interface for advanced SPECT reconstruction algorithms implemented on a remote massively parallel computer

Data from Single Photon Emission Computed Tomography (SPECT) studies are blurred by inevitable physical phenomena occurring during data acquisition. These errors may be compensated by means of reconstruction algorithms which take into account accurate physical models of the data acquisition procedure. Unfortunately, this approach involves high memory requirements as well as a high computational burden which cannot be afforded by the computer systems of SPECT acquisition devices. In this work the possibility of accessing High Performance Computing and Networking (HPCN) resources through a World Wide Web interface for the advanced reconstruction of SPECT data in a clinical environment was investigated. An iterative algorithm with an accurate model of the variable system response was ported on the Multiple Instruction Multiple Data (MIMD) parallel architecture of a Cray T3D massively parallel computer. The system was accessible even from low cost PC-based workstations through standard TCP/IP networking. A speedup factor of 148 was predicted by the benchmarks run on the Cray T3D. A complete brain study of 30 (64×64) slices was reconstructed from a set of 90 (64×64) projections with ten iterations of the conjugate gradients algorithm in 9 s which corresponds to an actual speed-up factor of 135. The technique was extended to a more accurate 3D modeling of the system response for a true 3D reconstruction of SPECT data; the reconstruction time of the same data set with this more accurate model was 5 min. This work demonstrates the possibility of exploiting remote HPCN resources from hospital sites by means of low cost workstations using standard communication protocols and an user-friendly WWW interface without particular problems for routine use.

Parallel Implementation of the PHOENIX Generalized Stellar Atmosphere Program

We describe the parallel implementation of our generalized stellar atmosphere and non-LTE (NLTE) radiative transfer computer program PHOENIX. We discuss the parallel algorithms we have developed for radiative transfer, spectral line opacity, and NLTE opacity and rate calculations. Our implementation uses a multiple instruction-multiple data design based on a relatively small number of MPI library calls. We report the results of test calculations on a number of different parallel computers and discuss the results of scalability tests.

Nonlinear finite element analysis of beams and arches using parallel processors

A nonlinear finite element formulation for beams and arches is presented for implementation on a multiple instruction multiple data (MIMD) parallel machine. A twelve degree of freedom arch element is developed using Green strains and 2nd Piola Kirchhoff stresses. Concepts like loop splitting and expression splitting are used to improve the load balancing capability among the processors. Using this developed element, beams and arches that are undergoing large displacements and large rotations are analyzed, and efficiency of the program is compared with a sequential implementation.

High-performance computing and networking as tools for accurate emission computed tomography reconstruction.

It is well known that the quantitative potential of emission computed tomography (ECT) relies on the ability to compensate for resolution, attenuation and scatter effects. Reconstruction algorithms which are able to take these effects into account are highly demanding in terms of computing resources. The reported work aimed to investigate the use of a parallel high-performance computing platform for ECT reconstruction taking into account an accurate model of the acquisition of single-photon emission tomographic (SPET) data. An iterative algorithm with an accurate model of the variable system response was ported on the MIMD (Multiple Instruction Multiple Data) parallel architecture of a 64-node Cray T3D massively parallel computer. The system was organized to make it easily accessible even from low-cost PC-based workstations through standard TCP/IP networking. A complete brain study of 30 (64x64) slices could be reconstructed from a set of 90 (64x64) projections with ten iterations of the conjugate gradients algorithm in 9 s, corresponding to an actual speed-up factor of 135. This work demonstrated the possibility of exploiting remote high-performance computing and networking resources from hospital sites by means of low-cost workstations using standard communication protocols without particular problems for routine use. The achievable speed-up factors allow the assessment of the clinical benefit of advanced reconstruction techniques which require a heavy computational burden for the compensation effects such as variable spatial resolution, scatter and attenuation. The possibility of using the same software on the same hardware platform with data acquired in different laboratories with various kinds of SPET instrumentation is appealing for software quality control and for the evaluation of the clinical impact of the reconstruction methods.

Steps towards steady-state simulation on MIMD machines: solving almost block diagonal linear systems

We study, in this paper, the feasibility of using preconditioned iterative methods to solve linear systems on MIMD (multiple instruction multiple data) machines when solving steady-state simulation problems. We show that, provided a good parallel preconditioner is available, the proposals in this paper give very good parallel performance. Our proposals are directly applicable to almost block diagonal linear systems, but can be extended to more general flowsheeting problems (Paloschi, 1996). This will be the subject of a future paper. A preconditioning strategy blending an incomplete LU decomposition with a polynomial preconditioner is proposed and tested with the iterative method GMRES of Saad and Schultz (1986). A parallel implementation is tested on two MIMD machines, a Fujitsu AP1000 with 128 processors and a cluster of workstations with PVM (Parallel Virtual Machine) (Geist et al., 1994). Two examples are used, a simple model of a distillation column using Antoine coefficients and the discretization of a PDE. The second example allows the variation in the size of the problem and results are presented for up to 65 536 equations on the AP1000 and 262 144 equations on the cluster. Almost perfect efficiency is observed when solving a problem of size 65 536 using 32 processors on the AP1000 (and size 16 384 using 16). Very good speed-ups also are obtained on the cluster of workstations (a speed-up of 15.3 using 16 processors) when solving a problem of size 65 536. The results on the cluster are presented for different load conditions on the network.

Parallel solvers for nonlinear elliptic problems based on domain decomposition ideas

In the present paper, the solution of nonlinear elliptic boundary value problems (b.v.p.) on parallel machines with Multiple Instruction Multiple Data (MIMD) architecture is discussed. Especially, we consider electro-magnetic field problems the numerical solution of which is based on finite element (FE) discretizations and a nested Newton solver. For solving the linear systems of algebraic FE equations in each Newton step, parallel conjugate gradient methods with a Domain Decomposition preconditioner (DD-PCG) as well as parallelized global multigrid (GMG) methods are applied. The implementation of the whole algorithm, i.e. the mesh generation, the generation of the FE equations, the nested Newton algorithm, the DD-PCG method and the GMG method, is based on a non-overlapping DD data structure. The efficiency of the parallel DD-PCG methods and the parallelized GMG methods, which are embedded in the nested Newton solver, are compared. Furthermore, the performance on different machines (GC Power Plus, Multicluster with transputers T805, and workstation cluster) is demonstrated by numerical results.

Efficient parallel Gauss-Doolittle matrix triangulation

An efficient parallel procedure for the triangulation of real symmetric (or complex Hermitian) matrices is presented. The methods of Gauss and Doolittle have previously been combined to produce a hybrid sequential method that was computationally faster than both. Numerical results demonstrate that this advantage is retained when MIMD (multiple instruction multiple data) distributed memory parallel processing is employed, so that parallel Gauss-Doolittle triangulation is faster than the equivalently parallelized Gauss elimination method.

PARALLEL IMPLEMENTATION FOR SOLVING THE TWO-DIMENSIONAL NAVIER-STOKES EQUATIONS ON A CELL-CENTERED, UNSTRUCTURED GRID

A compressible, two-dimensional, unstructured-grid Navier-Stokes solver has been developed for subsonic viscous flows. Pseudo-time preconditioning is applied to improve low Much number convergence for primitive as well as conserved variables. The scheme was implemented on a workstation and on the Ncube2s, which utilizes the multiple-instruction, multiple-data (MIMD) approach to parallel computing. A detailed description of the implementation procedure for an unstructured-grid flow solver on a MIMD computer is given. Test cases include steady as well as unsteady flows.

Parallel three-dimensional mesh generation on distributed memory MIMD computers

This paper discusses the development of an automatic mesh generation technique designed to operate effectively on multiple instruction multiple data (MIMD) parallel computers. The meshing approach is hierarchical, that is, model entities are meshed after their boundaries have been meshed. Focus is on the region meshing step. An octree is constructed to serve as a localization tool and for efficiency. The tree is also key to the efficient parallelization of the meshing process since it supports the distribution of load to processors. The parallel mesh generation procedure repartitions the domain to be meshed and applies on processor face removals until all face removals with local data have been performed. The portion of the domain to be meshed remaining is dynamically repartitioned at the octant level using an Inertial Recursive Bisection method and local face removals are reperformed. Migration of a terminal octant involves migration of the octant data and the octant's mesh faces and/or mesh regions. Results show relatively good speed-ups for parallel face removals on small numbers of processors. Once the three-dimensional mesh has been generated, mesh regions may be scattered across processors. Therefore, a final dynamic repartitioning step is applied at the region level to produce a partition ready for finite element analysis.

Parallel Algorithms for Orthotropic Problems

Finite element meshes and node-numberings suitable for parallel solution with equally loaded processors are presented for linear orthotropic elliptic partial differential equations. These problems are of great importance, for instance in the oil and airfoil industries. The linear systems of equations are solved by the conjugate gradient method preconditioned by modified incomplete factorization, MIC. The basic method presented, is based on fronts of uncoupled nodes and unlike earlier methods it has the advantage of no requirement of a specific orientation of the mesh. This method is however, in general, restricted to small degree of anisotropy in the differential equation. Another method, which does not suffer from this limitation, uses rotation of the differential equation and spectral equivalence. The rotation is made in such a way that in the new co-ordinate system, the basic method is applicable. The spectral equivalence property is used for estimation of the condition number of the preconditioned system. Both methods are suitable for implementation on parallel computers. The computer architecture could be single instruction multiple data (SIMD) as well as multiple instruction multiple data (MIMD) with shared or distributed memory. Implementation of the basic method on a shared memory parallel computer shows a significant improvement by use of the MIC method compared with the diagonal scaling preconditioning method.