Thinking Machines Research Articles

Evaluation of architectural support for global address-based communication in large-scale parallel machines

Large-scale parallel machines are incorporating increasingly sophisticated architectural support for user-level messaging and global memory access. We provide a systematic evaluation of a broad spectrum of current design alternatives based on our implementations of a global address language on the Thinking Machines CM-5, Intel Paragon, Meiko CS-2, Cray T3D, and Berkeley NOW. This evaluation includes a range of compilation strategies that make varying use of the network processor; each is optimized for the target architecture and the particular strategy. We analyze a family of interacting issues that determine the performance trade-offs in each implementation, quantify the resulting latency, overhead, and bandwidth of the global access operations, and demonstrate the effects on application performance.

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.

Data-parallel lower-upper relaxation method for the Navier-Stokes equations

The lower-upper symmetric Gauss-Seidel method is modified for the simulation of viscous flows on massively parallel computers. The resulting diagonal data-parallel lower-upper relaxation (DP-LUR) method is shown to have good convergence properties on many problems. However, the convergence rate decreases on the high cell aspect ratio grids required to simulate high Reynolds number flows. Therefore, the diagonal approximation is relaxed, and a full matrix version of the DP-LUR method is derived. The full matrix method retains the data-parallel properties of the original and reduces the sensitivity of the convergence rate to the aspect ratio of the computational grid. Both methods are implemented on the Thinking Machines CM-5, and a large fraction of the peak theoretical performance of the machine is obtained. The low memory use and high parallel efficiency of the methods make them attractive for large-scale simulation of viscous flows.

Parallelization of direct simulation Monte Carlo method combined with monotonic Lagrangian grid

The monotonic Lagrangian grid (MLG) and the direct simulation Monte Carlo (DSMC) methodology were combined on the Thinking Machines CM-5 to create a fast DSMC-MLG code with automatic grid adaptation based on local number densities. The MLG is a data structure in which particles that are close in physical space are also close in computer memory. Using the MLG data structure, physical space is divided into a number of templates (cells), each containing the same number of particles. An MLG-regularization method, stochastic grid restructuring, is implemented to minimize the occurrence of highly skewed cells. Parallelization of the DSMC-MLG is achieved by two different mapping techniques. First, simulated particles are mapped onto the parallel processors for the particle-oriented processes, such as convection, boundary interactions, and MLG sorting. Second, particle templates are mapped onto the processors for computing the macroscopic quantities (i.e., pressure, velocity, density, and temperature) and statistical sampling. In both levels of mapping, the code logic focuses on the structured and fast communications on the CM-5 architecture. The computing time required by the parallel DSMC-MLG code was significantly decreased compared with other parallel efforts and its parallel efficiency on 512 processors achieved approximately 80% for simulation involving one-half million particles.

Design of high level modelling / high performance simulation environments

Advances in massively parallel platforms are increasing the prospects for high performance discrete event simulation. Still the difficulty in parallel programming persists and there is increasing demand for high level support for building discrete event models to execute on such platforms. We present a parallel DEVS-based (Discrete Event System Specification) simulation environment that can execute on distributed memory multicomputer systems with benchmarking results of a class of high resolution, large scale ecosystem models. Underlying the environment is a parallel container class library for hiding the details of message passing technology while providing high level abstractions for hierarchical, modular DEVS models. The C++ implementation working on the Thinking Machines CM-5 demonstrates that the desire for high level modeling support need not be irreconcilable with sustained high performance.

Alternating direction splitting for block Angular parallel optimization

We develop and compare three decomposition algorithms derived from the method of alternating directions. They may be viewed as block Gauss-Seidel variants of augmented Lagrangian approaches that take advantage of block angular structure. From a parallel computation viewpoint, they are ideally suited to a data parallel environment. Numerical results for large-scale multicommodity flow problems are presented to demonstrate the effectiveness of these decomposition algorithmims on the Thinking Machines CM-5 parallel supercomputer relative to the widely-used serial optimization package MINOS 5.4.

ASYNCHRONOUS PARALLEL ALGORITHMS BASED ON DDM ON CM-5 (I) FOR SOLVING LINEAR ELLIPTIC PDE'S∗

This is the first in a series of papers about asynchronous parallel algorithms based on domain decomposition method (DDM) for solving partial differential equations on MIMD machines, especially on CM-5. These algorithms are all called Schwarz-Chaotic Over Relaxation (S-COR), the generalization of the Schwarz Alternating Method. In this paper, the linear elliptic partial differential equations is discussed. Numerical results are presented to demonstrate the effectiveness of these decomposition algorithms on the Thinking Machines CM-5 parallel supercomputer.

Computational optimization of finite difference methods on the CM5

Techniques used to optimize a finite-difference program on a Thinking Machines' CM5 parallel processing system are presented. These techniques are discussed within several categories: vector unit optimization, separation of communications and computations, and optimal data partitioning. A simplified model is employed to illustrate these concepts. The results of applying these techniques to a more complicated finite-difference calculation are also reported.

Efficient Iterative Methods Applied to the Solution of Transonic Flows

We investigate the use of an inexact Newton's method to solve the potential equations in the transonic regime. As a test case, we solve the two-dimensional steady transonic small disturbance equation. Approximate factorization/ADI techniques have traditionally been employed for implicit solutions of this nonlinear equation. Instead, we apply Newton's method using an exact analytical determination of the Jacobian with preconditioned conjugate gradient-like iterative solvers for solution of the linear systems in each Newton iteration. Two iterative solvers are tested; a block s-step version of the classical Orthomin(k) algorithm called orthogonal s-step Orthomin (OSOmin) and the well-known GMRES method. The preconditioner is a vectorizable and parallelizable version of incomplete LU (ILU) factorization. Efficiency of the Newton–Iterative method on vector and parallel computer architectures is the main issue addressed. In vectorized tests on a single processor of the Cray C-90, the performance of Newton–OSOmin is superior to Newton–GMRES and a more traditional monotone AF/ADI method (MAF) for a variety of transonic Mach numbers and mesh sizes. Newton–GMRES is superior to MAF for some cases. The parallel performance of the Newton method is also found to be very good on multiple processors of the Cray C-90 and on the massively parallel thinking machine CM-5, where very fast execution rates (up to 9 Gflops) are found for large problems.

Data-parallel total variation diminishing method for sonic boom calculations

Sonic boom predictions are shown for the near and midfield and comparisons are made with experimental data. The computations are performed on the Thinking Machines' CM-5 massively parallel supercomputer to utilize its large available memory and high floating point performance. A second-order-accurate total variation diminishing scheme is used to solve the Euler equations in the computations. Additionally, a recently developed implicit method, based on the LU-SGS algorithm, is used to speed the convergence and accuracy of the steady- state computations. The method is shown to work well on near- and midfield sonic boom predictions for several test cases. HE projected use of the high speed civil transport (HSCT) has drawn attention to the problems associated with the noise due to sonic booms.1 An accurate and efficient sonic boom prediction methodology is needed for the assessment of various proposed HSCT configurations. One such method that utilizes the power of massively parallel supercomputers is presented here. A review of current sonic boom prediction methods is given by Plotkin.2 Many of these methods are based on the modified linear analysis of Whitham 3 and Walkden.4 Other methods have been developed based on a modified method of char- acteristics that approximately account for the effects of three- dimensional flows. Experimental and analytical studies have shown that these methods lose effectiveness as freestream Mach number approaches 3. In high Mach number regions there are strong shock waves with significant entropy gen- eration. The linear-based methods neglect these production terms.5 A modified method of characteristi cs has been de- veloped that can account for these terms,6 but it requires nonlinear near-field initial data that is difficult to obtain com- putationally or experimentall y. Several authors have developed prediction methods based on near-field solutions of the Euler or Navier-Stokes equa- tions.7-8 Most of these methods involve marching in one spatial dimension with an implicit Euler scheme or a parabolized Navier-Stokes (PNS) algorithm. These solutions are used in conjuction with Whitham's F-function to evaluate the far-field pressure signature. The existing prediction methods are un- attractive for several reasons. First, additional accuracy in the marching direction greatly increases the computational time. Also, they do not employ methods designed to resolve shock waves effectively. Finally, they are not easily implemented on massively parallel supercomputers. Accurate shock resolution is needed in the computational fluid dynamic (CFD) methods to determine the far-field pres- sure signature effectively. Since the prediction methods are dependent on shock resolution it is desirable to use a com- putational method that is designed to capture shock waves. The total variation diminishing (TVD) schemes developed by Harten,9 Yee,10 and others are well suited for this purpose. In this study a second-order-accurate TVD scheme is used to determine the pressure field about several models. These so- lutions are then compared with experiment.

Philosophy of Technology in Spanish Speaking Countries

Edesao de la tecnolo Authors and Issues in Philosophy of Technology in the Spanish-Speaking World, C. Mitcham. Part I: From Chile - Technology and Politics: Toward Artificial History? M. Gar de la Huerta Globalization: Homogenization with and Increasing Gap, M. Gara de la Huerta Thinking Machines and the Crisis of Modern Reason, E. Sabrovsky Ethics, Science, and Technology, C. Verdugo. Part II: From Costa Rica - Contributions to the Philosophy of Technology in Costa Rica, L.A. Camacho Naranjo Science, Technology, and Development - Some Models of their Relationship, L.A. Camacho Naranjo Ethics, Pernicious Technology, and the Technological Argument, E.R. Rarez Brico. Part III: From Mexico - Technology and Basic Needs: Proposal for a Debate on Fundamental Criteria, E. Dussel On Technique, J. Gaos Objects and their Epistemological Base, H. Padilla Satellites and our Morality, L. Zea. Part IV: From Spain - Technology and Human Nature, M.L. Gara-Merita Philosophy, Technology, and Society, M. Medina Does Technology Construct Scientifc Reality?, R. Queral Moreno The Design and Evaluation of Technologies: Some Conceptual Issues, M.A. Quintanilla From World to the Social Assessment of Technology: Remarks on Science, Technology, and Society, J. Sanmarn Genethics: the Social Assessment of the Risks and Impacts of Genetic Engineering, J. Sanmarn. Part V: From Venezuela - Science, Technology, History, and Philosophy in the Cultural Atmosphere of our Time, J.D. Gara Bacca The Present and Future of Humanity, E. May Vallenilla Education for Freedom versus Socio-Technical Control by Pedagogical Means, L. Molina. Part VI: From Other Americas - Two Philosophical Approaches to the Problem of Technics and their Meaning for Latin America, R. Fornet-Betancourt Science, Technology, and Society Education in the Latin American Context, M.M. Borrero The Moral Vision of Technology in Contemporary Latin American Fiction, J. Robinett The Social Implications of Information Technologies - a Latin American Perspective, J. Sutz.

PARAMICS?Parallel microscopic simulation of road traffic

This paper describes work done on the original PARAMICS project, which was developed for the Edinburgh Parallel Computing Centre to examine parallel microscopic road traffic simulation. The simulator, constructed originally for a Thinking Machines Connection Machine (CM-200), uses a data-parallel approach to simulate approximately 200,000 vehicles on 20,000 miles of roadway. More recent work has focused on the use of a message-passing paradigm, with a 256-node CRAY T3D as the target machine. The message-passing version of PARAMICS, PARAMICS-MP, is inherently scalable and can model many smaller networks on a broad range of platforms.

Practical parallel algorithms for personalized communication and integer sorting

A fundamental challenge for parallel computing is to obtain high-level, architecture independent, algorithms which efficiently execute on general-purpose parallel machines. With the emergence of message passing standards such as MPI, it has become easier to design efficient and portable parallel algorithms by making use of these communication primitives. While existing primitives allow an assortment of collective communication routines, they do not handle an important communication event when most or all processors have non-uniformly sized personalized messages to exchange with each other. We focus in this paper on the h-relation personalized communication whose efficient implementation will allow high performance implementations of a large class of algorithms. While most previous h-relation algorithms use randomization, this paper presents a new deterministic approach for h-relation personalized communication with asymptotically optimal complexity for h>p 2 . As an application, we present an efficient algorithm for stable integer sorting. The algorithms presented in this paper have been coded in Split-C and run on a variety of platforms, including the Thinking Machines CM-5, IBM SP-1 and SP-2, Cray Research T3D, Meiko Scientific CS-2, and the Intel Paragon. Our experimental results are consistent with the theoretical analysis and illustrate the scalability and efficiency of our algorithms across different platforms. In fact, they seem to outperform all similar algorithms known to the authors on these platforms.

File-access characteristics of parallel scientific workloads

Phenomenal improvements in the computational performance of multiprocessors have not been matched by comparable gains in I/O system performance. This imbalance has resulted in I/O becoming a significant bottleneck for many scientific applications. One key to overcoming this bottleneck is improving the performance of multiprocessor file systems. The design of a high-performance multiprocessor file system requires a comprehensive understanding of the expected workload. Unfortunately, until recently, no general workload studies of multiprocessor file systems have been conducted. The goal of the CHARISMA project was to remedy this problem by characterizing the behavior of several production workloads, on different machines, at the level of individual reads and writes. The first set of results from the CHARISMA project describe the workloads observed on an Intel iPSC/860 and a Thinking Machines CM-5. This paper is intended to compare and contrast these two workloads for an understanding of their essential similarities and differences, isolating common trends and platform-dependent variances. Using this comparison, we are able to gain more insight into the general principles that should guide multiprocessor file-system design.

Massively parallel implementation of a 3D vortex-boundary element method

Vortex-boundary element simulation of three-dimensional wall-bounded flows is performed on a massively parallel architecture using the data parallel paradigm. With proper optimization, implementation on the Thinking Machines CM5 is shown to yield over an order of magnitude speed-up over the Cray C90, using 128 processors. This makes the direct evaluation of the flow eld represented by up to 100 000 vortex elements feasible. In this paper, the CPU time and memory requirements for the direct evaluation of the vortical eld using three parallelization algorithms are compared. In addition, performance results for the evaluation of the vortical and potential velocities, and their gradients, are presented as a function of the number of vortex elements and the number of processing nodes. Finally, to demonstrate the capability of the developed method, preliminary results from the case of flow around a stationary, idealized trailer truck near the ground level at Re = 500 are presented.

Parallel algorithms for image enhancement and segmentation by region growing, with an experimental study

This paper presents efficient and portable implementations of a powerful image enhancement process, the Symmetric Neighborhood Filter (SNF), and an image segmentation technique that makes use of the SNF and a variant of the conventional connected components algorithm which we call δ-Connected Components. We use efficient techniques for distributing and coalescing data as well as efficient combinations of task and data parallelism. The image segmentation algorithm makes use of an efficient connected components algorithm based on a novel approach for parallel merging. The algorithms have been coded in Split-C and run on a variety of platforms, including the Thinking Machines CM-5, IBM SP-1 and SP-2, Cray Research T3D, Meiko Scientific CS-2, Intel Paragon, and workstation clusters. Our experimental results are consistent with the theoretical analysis (and provide the best known execution times for segmentation, even when compared with machine-specific implementations). Our test data include difficult images from the Landsat Thematic Mapper (TM) satellite data.

Parallel solution of large‐scale differential‐algebraic systems

AbstractDASPK solves large‐scale systems of differential‐algebraic equations. It is based on the integration method in DASSL, but instead of a direct method for the associated linear systems which arise at each time step, the preconditioned GMRES iteration is applied in combination with an inexact Newton method. Two parallel versions of DASPK have been developed: DASPKF90, a Fortran 90 data parallel implementation, and DASPKMP, a message‐passing implementation written in Fortran 77 with extended BLAS. The parallel versions have been implemented for the Thinking Machines Corporation (TMC) CM‐5, a massively parallel multiprocessor, keeping the user interface relatively simple while allowing for portability to other massively parallel architectures. The codes have been demonstrated on several large‐scale test problems, including three‐dimensional formulations of the heat equation, the Cahn‐Hilliard equation and a multi‐species reaction‐diffusion problem. The formulations are described, including detail on preconditioning the Krylov iteration, timing results and performance analysis.

Complete exchange on the CM-5 and Touchstone Delta

The complete exchange (or all-to-all personalized) communication pattern occurs frequently in many important parallel computing applications. It is the densest form of communication because all processors need to communicate with all other processors. This can result in severe link contention and degrade performance considerably. Hence, it is necessary to use efficient algorithms in order to get good performance over a wide range of message and multiprocessor sizes. In this paper we present several algorithms to perform complete exchange on the Thinking Machines CM-5 and the Intel Touchstone Delta multiprocessors. Since these machines have different architectures and communication capabilities, different algorithms are needed to get the best performance on each of them. We present four algorithms for the CM-5 and six algorithms for the Delta. Complete exchange algorithms generally assume that the number of processors is a power of two. However, on the Delta the number of processors allocated by a user need not be a power of two. We propose algorithms that are even applicable to non-power-of-two meshes on the Delta. We have developed analytical models to estimate the performance of the algorithms on the basis of system parameters. Performance results on the CM-5 and Delta are also presented and analyzed.

Numerical Simulation of Shock-Cylinder Interactions: I. Resolution

We apply two different high-order shock capturing schemes to the study of a two-dimensional unsteady inviscid flow. In particular, we study the interaction of a planar shock with o cylindrical volume of a light gas (helium or hydrogen) contained in air. The two schemes used are the Chebyshev collocation method and the ENO finite difference scheme of Osher and Shu, and they are applied to a physical model consisting of the Euler equations with a real gas equation of state and multiple chemical species. The parallel implementation and low-level coding of the ENO scheme on the Thinking Machines CM-5 results in much higher performance than is possible on a standard serial or vector machine. The ENO code is compared with an existing experimental result and agrees well with it. The results of spectral and ENO calculations are then compared with each other at different resolutions for a Mach 2 interaction. The spectral scheme, though highly oscillatory in nature for discontinuous problems (Gibbs), accurately predicts both large and fine scale structures of the interaction between the shock and the light gas cylinder. Good results can be recovered from the spectral results by post-processing the raw numerical data to remove the Gibbs phenomena. These results are compared with the ENO schemes. The comparison is progressively better as the grid refinement and numerical order of the ENO scheme is increased. This demonstrates definitively the applicability and value of high order schemes to flows with shocks and complicated non-linear physics.

What if computers could think?

The question whether or not computers can think was first asked in print by Alan Turing in his seminal 1950 article. In order to avoid defining what a computer is or what thinking is, Turing resorts to “the imitation game” which is a test that allows us to determine whether or not a machine can think. That is, if an interrogator is unable to tell whether responses to his questions come from a human being or from a machine, the machine is imitating a human being so well that it has to be acknowledged that these responses result from its thinking. However, then as now, it is not an indisputable claim that machines could think, and an unceasing stream of papers discussing the validity of the test proves this point. There are many arguments in favour of, as well as against, the claims borne by the test, and Turing himself discusses some of them. In his view, there are mice possible objections to the concept of a thinking machine, which he eventually dismisses as weak, irrelevant, or plain false. However, as he admits, he can present “no very convincing arguments of a positive nature to support my views. If I had I should not have taken such pains to point out the fallacies in contrary views”.