Multiprocessor Machines Research Articles

BSP linear solvers for dense matrices

This paper presents a study of four parallel linear system solvers implemented using the Bulk Synchronous Parallel programming model. We show that better performances are achieved by the implicit methods due both to higher exploitation of hardware capabilities and to the lower communication costs. A BSP performance model of all the algorithms has been built and the predictions obtained have been compared with the results of measurements on two different parallel systems: a cluster of workstations and a shared memory multiprocessor machine. The implicit algorithms have proved to be from 20% to 40% faster than their traditional counterparts. The high accuracy of prediction makes the BSP programming paradigm an ideal framework for developing parallel algorithms.

Linearized Formulation of the Riemann Problem for Adiabatic and Isothermal Magnetohydrodynamics

In this paper we present a linearized formulation of the Riemann problem for ideal adiabatic and isothermal magnetohydrodynamic (MHD). Roe's property U is obtained. This ensures good capturing of shocks of arbitrary strength. A parameter vector is found which gives an almost consistent mean state, where consistency is defined in the sense of Roe. The eigenvalues and eigenvectors for the linearized Riemann problem for MHD have been explicitly constructed further facilitating numerical implementation. The method has been tested in the author's RIEMANN code for numerical MHD. In a companion paper on TVD methods for MHD, where this Riemann solver was used, we have shown that excellent results can be obtained. The consistency for this formulation is illustrated by the fact that the eigenvalues have the same symmetrical distribution around the Roe-averaged x-component of the velocity that the eigenvalues for ideal MHD have around the physical x-component of the velocity. The accuracy, simplicity, and noniterative nature of the linearized formulation of the MHD Riemann problem presented here make it especially suitable for use on massively parallel processor machines.

Transfer of Fixation Using Affine Structure: Extending theAnalysis to Stereo

This paper describes an algorithm for maintaining fixation upon a 3D body-centred point using 3D affine transfer, extending an earlier monocular method to stereo cameras. Transfer is based on corners detected in the image and matched over time and in stereo. The paper presents a method using all available matched data, providing immunity to noise and poor conditioning. The algorithm, implemented at video rates on a multi-processor machine, incorporates controlled degradation in the presence of insufficient data. Results are given from experiments using a four-axis active stereo camera platform, first which show the greater stability of the fixation point over the monocular method, both as it appears in the image and occurs in the scene; and, secondly, which show the recovery and evolution of 3D affine structure during fixation. It is shown that fixation and explicit structure recovery can occur separately, allowing the information required for gaze control to be computed in a fixed time.

Processor—Time Tradeoffs under Bounded-Speed Message Propagation: Part I, Upper Bounds

Upper bounds are derived for the processor-time tradeoffs of machines such as linear arrays and two-dimensional meshes, which are compatible with the physical limitation expressed by bounded-speed propagation of messages (due to the finiteness of the speed of light). It is shown that parallelism and locality combined may yield speedups superlinear in the number of processors. The speedups are inherent, due to the optimality of the obtained tradeoffs as established in a companion paper. Simulations are developed of multiprocessor machines by analogous machines with fewer processors. A crucial role is played by the hierarchical nature of the memory system. A divide-and-conquer technique for hierarchical memories is developed, based on the graph-theoretic notion of topological separator. For multiprocessors, this technique also requires a careful balance of memory access and interprocessor communication costs, which leads to non-intuitive orchestrations of the simulation process.

Role of parallelism in wavelet analysis of radar imaging

The continuous wavelet analysis method developed by J. Bertrand, P. Bertrand and J. P. Ovarlez [14] represents target reflectivity by a function on a space with four dimensions, commonly called radar hyperimage. It represents a theoretical formulation of radar imaging that is an improvement and an extension of the tomography technique developed by C. Pouit [2] in the 1970s. The hyperimage is becoming a new tool for perceiving microwaves, an interface between radar and the eye, providing the representation of the information contained in all the radar measurements, also called hologram. The hyperimage is easily constructed by an efficient, original algorithm that uses fast computation tools based on fast Fourier transforms (FFT). The computation load is however large and the execution time required by the method on a sequential computer is prohibitive. However, the structure of the algorithm naturally gives the idea of implementing it on a multiprocessor machine. Our work shows that programming the algorithm on a parallel computer makes it possible to approach real time performance and make the software interactive. In effect, the performance of our HYPIM programme developed on ONERA's iPSC860 has demonstrated that the combined use of the many facilities offered by wavelet analysis and by parallel computation improves the capabilities of radar imaging.

Exploiting IEEE 1355 Routable Serial Links in a Real-Time Vision Architecture

A multi-processor machine vision system typically uses several different interconnection technologies to satisfy its various communication needs. This paper describes a machine vision architecture which uses high-speed packet-switched serial links (based on Inmos DS-links and conforming to the IEEE 1355 standard) as itssolecommunication technology. This technology offers significant advantages in terms of configurability, modularity, and scalability. It leads to an efficient ‘building block’ approach to constructing tightly coupled vision machines for specific applications from a general purpose set of modules. We are building a demonstrator system combining T9000 transputers with our own hardware modules which perform image I/O, correlation, and convolution; and message multi-casting modules. All modules conform to the industry standard HTRAM format. The paper discusses the implementation issues surrounding the use of the T9000, C104, and C101 DS-link components, and shows how they may be used to integrate hardware and software processing modules into a single conceptual model.

Experience with the design of a performance tuning tool for parallel programs

Designing parallel programs to maximise the utilisation of the processing power of a multiprocessor machine requires a series of performance analysis and tuning. This paper describes a performance tuning tool for message-passing parallel programs. The tool uses a logical clock approach to instrument the program for collecting performance data. Combining the advantages of relational databases and three-dimensional spreadsheets to organise the performance data, the tool provides visual assistance for analysing the program performance. Various graphical displays which help the user to fine-tune the performance of message-passing programs are discussed.

Task assignment in Cayley interconnection topologies

Based on the Cayley graph framework for the generation and evaluation of multiprocessor interconnection topologies, an application dependent task mapping problem is addressed. We start with interconnection topologies considered attractive from a graph theoretical point of view. Given such a topology together with an application dependent task communication profile, the problem addressed in this paper is to find an optimal task-to-processor assignment. Such a mapping yields for the given interconnection topology and the given profile a minimal expected communication path length and therefore a minimal number of data transfer steps between the physical processing elements of a multiprocessor machine. At first, the Cayley graph approach is briefly outlined. We demonstrate the potential of Cayley graphs for a systematic generation framework aiming at node-symmetric interconnection topologies for a given number of processing elements each equipped with a constant number of communication channels. Cayley graphs found most attractive within a large set of generated graphs are compared to prominent interconnection topologies like, for example, hypercubes. Later on, it is shown that the problem of the optimal task-to-processor assignment, being a special case of the well-known quadratic assignment problem, is still NP-hard. Consequently, any practically relevant mapping algorithm can be expected to produce at best near-optimal solutions for reasonable problem sizes beyond approximately 10 nodes. We present two fundamentally different mapping approaches, namely a straightforward greedy mapping and a more sophisticated algorithm using simulated annealing techniques as known from artificial intelligence applications. For both approaches, we elaborate on their relative performance as well as, where feasible, on the question of suboptimality.

Nonintrusive Cloning Garbage Collection with Stock Operating System Support

It is well accepted that automatic garbage collection simplifies programming, promotes modularity, and reduces development effort. However it is commonly believed that these advantages do not counteract the perceived price: excessive overheads, possible long pause times while garbage collections occur, and the need to modify existing code. Even though there are publically available garbage collector implementations that can be used in existing programs, they do not guarantee short pauses, and some modification of the application using them is still required. In this paper we describe a snapshot-at-beginning concurrent garbage collector algorithm and its implementation. This algorithm guarantees short pauses, and can be easily implemented on stock UNIX-like operating systems. Our results show that our collector performs comparable to other garbage collection implementations on uniprocessor machines and outperforms similar collectors on multiprocessor machines. We also show our collector to be competitive in performance with explicit deallocation. Our collector has the added advantage of being non-intrusive. Using a dynamic linking technique and effective root set inferencing, we have been able to successfully run our collector even in commercial programs where only the binary executable and no source code is available. In this paper we describe our algorithm, its implementation, and provide both an algorithmic and a performance comparison between our collector and other similar garbage collectors. ©1997 by John Wiley & Sons, Ltd.

New computing systems and future high-performance computing environment and their impact on structural analysis and design

Recent advances in computer technology that are likely to impact structural analysis and design are reviewed. A brief summary is given of the advances in microelectronics and networking technologies, and in the human-computer interaction paradigms and techniques. The major features of new and projected computing systems, including high-performance computers, parallel processing machines, and small systems are described. Advances in programming environment, numerical algorithms, and computational strategies for new computing systems are reviewed. The impact of the advances in computer technology on structural analysis and design is described. A scenario for future computing paradigm is presented and the near-term needs in the computational structures area are outlined.

Robust Parallel Algorithms for Solution of Riccati Equation

Robust and efficient parallel-vector algorithms are presented for the solution of the Riccati equations encountered in optimal control problems on shared-memory multiprocessor machines. The algorithms have been implemented on a Cray YMP 8/8128 and applied to three large problems resulting from a continuous bridge structure, a 21-story space truss structure, and a 12-story space moment-resisting building structure. Efficiency of the algorithms is presented in terms of millions of floating point operations per second (MFLOPS) and the speedup. The MFLOPS for the largest example resulting from the 12-story space frame structure is a high 206.0. The speedup due to parallel processing only (for the same example) using seven processors, is 6.33. When vectorization is combined with parallel processing a very significant speedup of 54.4 is obtained using seven processors. The algorithms developed in this research find applications in the complex integrated control/structural optimization problem. Further, the writers are currently using them to develop large adaptive/smart structures.

Design and implementation of parallel time-dependent least time path algorithms for Intelligent Transportation Systems applications

The development of Intelligent Transportation Systems (ITS) and the resulting need for real-time traffic management and route guidance models require fast shortest-path algorithms that can account for the dynamics of traffic networks. The objective of this paper is to introduce parallel designs for time-dependent shortest-path algorithms that can be used in real-time ITS applications. In this paper, two shared-memory and one message-passing algorithms are designed, implemented, coded and computationally tested on actual and random networks. The reported tests are performed on CRAY supercomputers, but the algorithms can be readily ported to lower-end multiprocessor machines.

Parallelism at CERN: real-time and off-line applications in the GP-MIMD2 project

A wide range of General Purpose High-Energy Physics applications, ranging from Monte carlo simulation to data acquisition, from interactive data analysis to on-line filtering, have been ported, or developed, and run in parallel on IBM SP-2 and Meiko CS-2 CERN large multi-processor machines. The ESPRIT project GP-MIMD2 has been a catalyst for the interest in parallel computing at CERN. The project provided the 128 processors Meiko CS-2 system that is now succesfully integrated in the CERN computing environment. The CERN experiment NA48 was involved in the GP-MIMD2 project since the beginning. NA48 physicists run, as part of their day-to-day work, simulation and analysis programs parallelized using the Message Passing Interface MPI. The CS-2 is also a vital component of the experiment Data Acquisition System and will be used to calibrate in real-time the 13 000 channels liquid krypton calorimeter.

Solution-Domain-Decomposition Method for Heat Transfer Problem Using Parallel Distributed Computing

Solution-domain-decomposition (SDD) method is formulated for solving heat transfer problem and generalized for solving multi-domain problem. A generalized algorithm is suggested for parallel and distributing computation. Chebyshev expansion on the dependent variables is used for pseudospectral approximation of the governing equation in this study. Linear superposition principle is adapted to incorporate the interactions between the subdomains. By effective subdivision of computational domain, significant computational efficiency and computational memory savings are accomplished without losing spectral accuracy of the solution. Owing to independent characteristics of the subdomains. the scheme is well suited for multi-processor machines. Convergence study reveals that spectra! accuracy is still conserved for the multi-domain calculation. The calculation domain is divided up to 8 subdomains and calculation is distributed up to independent CPUs. Significant speed-up ratio is obtained by distributing the subtasks through the network.

Multiprocessing for neutron diffusion and deterministic transport methods

Multiprocessing is the most exciting development in computer technology in recent years, opening the door for enormous computational power limited only by the imagination and patience of the hardware developer, and the wit and perseverance of the algorithm designer. The potential benefit of this innovation was immediately recognized by researchers in the nuclear field, among others, who explored and reported their experience with concurrent implementation of algorithms, standard and new, to solve neutron diffusion and transport problems. The accumulated experience in, as well as the current status of, this rapidly changing area is reviewed. Collectively these illustrate the strong coupling between parallel algorithms and their target architectures to the extent that standard sequential schemes in nuclear applications are best suited to coarse grained, e.g. shared memory supercomputers, and medium grained, e.g. a few hundred networked powerful CPUs, platforms. The full potential of parallel computing will be better realized by novel, non-traditional, solution algorithms that best utilize all the components of a multiprocessor machine.

List Ranking and List Scan on the CRAYC90

Although parallel algorithms using linked lists, trees, and graphs have been studied extensively by the research community, implementations have met with limited success, even for the simplest algorithms. In this paper we present our results of a careful implementation study of parallel list ranking (and the related list-scan operation) and show that it can have substantial speed up over fast workstations. Obtaining good parallel performance for list ranking is a challenge for two reasons. First, although there are many asymptotic work-efficient algorithms, it is hard to keep the constants comparable to those of the sequential algorithm. In this paper we introduce a new parallel algorithm that both is work efficient and has small constant but sacrifices logarithmic time; it achieves only anO(log2n) time. We contend, however, that work efficiency and small constants are more important, given that multiprocessor machines are used for problems that are much larger than the number of processors and, therefore, the optimalO(logn) time is never achieved in practice. Second, list ranking is highly communication intensive and its memory access patterns are dynamic. We show, however, that by using high memory bandwidth multiprocessors, such as CrayC90 computers, programmed with virtual processors to hide latency we can ameliorate these problems. To the best of our knowledge, our implementation of list ranking and list scan on the CrayC90 is the fastest to date and is the first implementation that substantially outperforms fast workstations. The success of our algorithm is due to its moderate grain size and simplicity; the success of the implementation is due to pipelining reads and writes through vectorization to hide latency and optimizing performance by analyzing the expected execution time of the algorithm.

Data parallel computation for thermo-viscoelastic analysis of composite structures

Data parallel implementation of the thermo-viscoelastic finite element algorithm with a conjugate gradient (CG) method on the CM-5 is presented. The performance study of thermo-viscoelastic finite element procedures on massively parallel processing machines is conducted. Parametric studies on the CM-5 as well as the CRAY Y-MP and benchmarks for several problem sizes are shown. The performance of the conjugate gradient method on the CRAY Y-MP and CM-5 is evaluated and is compared with ones of the CRAY sparse matrix and the Feable solvers. Using the nine-node isoparametric shell element and 256-processor CM-5, a speed-up factor of about 18 over the CRAY Y-MP was obtained for calculating 160000 element stiffness matrices at a rate of 3·29 Gflops. As for the CG method, the CM-5 out-performed the CRAY Y-MP by a factor of 7 for solving 320 800 equations.

Formulation of methods for real-time simulation of biomechanical vocal tract

Based on a method that is related to the mortar finite element method, a domain subdivision method was developed which can be used towards the goal of a real-time simulation of the vocal tract. Previous work [R. Wilhelms-Tricarico, J. Acoust. Soc. Am. 97, 3085–3098 (1995)] on simulation of soft tissue for speech production research has resulted in rather slow algorithms that are not very suitable for implementation in parallel multiprocessor machines. A new formulation uses finite elements with quadratic interpolation of the displacement and velocity fields as well as four pressure variables per element, to approximate tissue incompressibility. This formulation is combined with a domain subdivision method that is designed to implement the simulation algorithm on a multi-processor computer such as CM5. The interaction between the various connected domains is achieved by an augmented Lagrangian method. [Work supported by NIH.]

A Partitioning-Independent Paradigm for Nested Data Parallelism

A generalization of the data parallel model has been proposed by Blelloch which permits the nesting of data parallel operators to specify parallel computation across nested and irregular data structures. In this paper we consider the costs of supporting the general model of nested data parallelism, analyzing the requirements such a model places upon an underlying model of execution. We propose a new multi-node execution model which meets the needs of the paradigm and is additionally generic in the partitioning of data aggregates within the system. The basis for our execution model is an abstract machine based upon elementary notions of nodal multi-threading. We demonstrate the utility of our proposal by providing a full definition for a simple nestable one-dimensional data parallel operator. We discuss the applicability of our design to existing multi-processor machines, illustrating performance statistics gathered from a prototype system we have constructed on the Thinking Machines CM-5.

Simulation of Synchrotron White-Beam Topographs. An Algorithm for Parallel Processing: Application to the Study of Piezoelectric Devices

This paper presents a new algorithm for the integration of Takagi–Taupin equations taking into account the fact that X-ray diffraction is a parallel phenomenon. The diffraction equations show that the propagation of the waves is independent in each incidence plane. It is thus possible to compute in parallel the propagation of the waves in different planes. Two algorithms are presented: the first one for multiprocessor machines where the processors share a common memory, the second one for massively parallel computers. The program is written to achieve a high vectorization ratio and to make it as efficient as possible with modem superscalar and array processors. The simulation of the image of a defect has been divided into two independent parts. In the first one, one computes the derivatives of the deformation inside the crystal; in the second one, these results are used to simulate the image. This allows one rapidly to change the model for a defect, something that was not feasible in all previously written simulation programs since the computation of the deformation was part of the simulation. The study of stroboscopic images of the propagation of acoustic waves in piezoelectric devices is given as an example of the possibilities of this new program.