High Performance Computing Systems and Applications contains the fully refereed papers from the 13th Annual Symposium on High Performance Computing, held in Kingston, Canada, in June 1999. This book presents the latest research in HPC architectures, distributed and shared memory performance, algorithms and solvers, with special sessions on atmospheric science, computational chemistry and physics. High Performance Computing Systems and Applications is suitable as a secondary text for graduate level courses, and as a reference for researchers and practitioners in industry.
This paper is devoted to the improvement of a numerical Maxwell-Schrödinger system, modeling the intense, high frequency and ultrashort laser-gas interaction and propagation, and previously presented in [3]. To this original model is added an equation of evolution of the free electron density, allowing then to describe precisely the current density in the Maxwell equations. Free electrons are obtained by absorption of the wavefunctions at the boundary of their corresponding TDSE computational domain. As a consequence higher intensity and longer pulses can then be considered using this new model compared to [3]. The new computing strategy is then presented, following [4].
The HPC world is waking up to the importance of storage as an integral part of an HPC solution. Many HPC sites are realizing that HPC class storage coupled with classic enterprise features such as snapshots, replication, disaster recovery, and ILM (Integrated Life Cycle Management), can result in the maximum performance and uptime for HPC applications. However, the biggest problem that HPC centers face is that they don't understand how their applications perform IO (i.e. IO pattern). This means that they can't define meaningful requirements for HPC storage. Knowledge of the IO patterns allows better definition of requirements including both time and money. This paper presents a technique for obtaining IO information and analyzing it to present IO pattern information. The technique is independent of the application.
With the increasing popularity of cloud computing, Hadoop has become a widely used open source cloud computing framework for large scale data processing. However, few efforts have been made to demonstrate the applicability of Hadoop to various real-world application scenarios in fields other than server side computations such as web indexing, etc. In this paper, we use the Hadoop cloud computing framework to develop a user application that allows processing of scientific data on clouds. A simple extension to Hadoop’s MapReduce is described which allows it to handle scientific data processing problems with arbitrary input formats and explicit control over how the input is split. This approach is used to develop a Hadoop-based cloud computing application that processes sequences of microscope images of live cells, and we test its performance. It is discussed how the approach can be generalized to more complicated scientific data processing problems.
Due to the increased use of composite materials in the aerospace industry, numerous attempts have been made to develop fatigue models in order to predict the fatigue behaviour and consequently the fatigue life of these materials. Existing fatigue models have significant deficiencies, thus are not widely acceptable in the industry. A better understanding of the exhibited fatigue behaviour of composite materials is consequently required. The complex nature of fatigue behaviour in fiber-reinforced composite materials is presently investigated. An explicit progressive damage model, that is mechanistic in nature, is currently being developed using the concept of a representative volume element. A micromechanical finite element model that is capable of explicit damage initiation and propagation modeling is utilized for simulation of damage development. The predicted numerical results illustrate the capabilities of the current model. Future work is also outlined in the paper as the development of the fatigue model is continued.
A parallel iterative algorithm is described for efficient solution of the Schur complement (interface) problem arising in the domain decomposition of stochastic partial differential equations (SPDEs) recently introduced in [1,2]. The iterative solver avoids the explicit construction of both local and global Schur complement matrices. An analog of Neumann-Neumann domain decomposition preconditioner is introduced for SPDEs. For efficient memory usage and minimum floating point operation, the numerical implementation of the algorithm exploits the multi-level sparsity structure of the coefficient matrix of the stochastic system. The algorithm is implemented using PETSc parallel libraries. Parallel graph partitioning tool ParMETIS is used for optimal decomposition of the finite element mesh for load balancing and minimum interprocessor communication. For numerical demonstration, a two dimensional elliptic SPDE with non-Gaussian random coefficients is tackled. The strong and weak scalability of the algorithm is investigated using Linux cluster.
Steady (RANS) and unsteady (URANS) numerical simulations of three-dimensional transonic flow in a high-pressure turbine have been conducted. The results are compared by evaluating differences in contours of total pressure loss, contours of static pressure and wakes. Significant differences have been observed in these parameters, both qualitative and quantitative, which demonstrate that URANS simulations are necessary when the effect of stator-rotor interaction on flows in turbines is strong.
We present an accurate method for the nonperturbative numerical solution of the Time-Dependent Schroedinger Equation, TDSE, for molecules in intense laser fields, using cylindrical/polar coordinates systems. For cylindrical coordinates systems, after use of a split-operator method which separates the z direction propagation and the (x,y) plane propagation, we approximate the wave function in each (x,y) section by a Fourier series ( ∑ c_m(ρ)e^imϕ ) which offers an exponential convergence in the φ direction and is naturally applicable for polar coordinates systems. The coefficients (c m (ρ)) are then calculated by a Finite Difference Method (FDM) in the ρ direction. We adopt the Crank-Nicholson method for the temporal propagation. The final linear system consists of a set of independent one-dimensional (ρ) linear systems and the matrix for every one-dimensional linear systems is sparse, so the whole linear system may be very efficiently solved. We note that the Laplacean operator in polar/cylindrical coordinate has a singular term near the origin. We present a method to improve the numerical stability of the Cranck-Nicholson method in this case. We illustrate the improved stability by calculating several eigenstates of H_3^++ with propagation in imaginary time. Two methods of spatial discretization are also compared in calculations of Molecular High Order Harmonic Generation, MHOHG.
High-performance computing is often used in various fields of computational science and engineering where the availability of a large memory is crucial to represent given problems. For problems arising from the discretization of partial differential equations, adaptivity is an alternative approach to reduce the storage requirements. We present an illustrating example demonstrating the need for an algorithm change when a serial hierarchical data structure is parallelized. More precisely, we investigate this issue taking as example an algorithm to transform finite element functions defined on an adaptive grid into functions on a Cartesian grid. By combining distributed- and shared-memory parallelization, the resulting hybrid parallel algorithm involves two levels of parallelism. The performance of the parallel algorithm is shown to be highly problem dependent. There are certain problem instances where the speedup is satisfactory while there is a dramatic performance reduction for other problem instances. The numerical experiments are given for transformations resulting from three-dimensional computational fluid dynamic problems. The problem instances are given by up to 750 000 tetrahedra representing about 3 millions of unknown variables. The experiments are carried out on a Xeon-based cluster using up to 64 processors.
We discuss a new methodology for implementing polarizable and flexible molecular models - the fluctuating charge and intramolecular potential (fCINTRA) method - in Molecular Dynamics (MD) simulations. An example has been provided for ethanol. In these models, all potential parameters depend on the local electrostatic field generated by the other molecules in the system. A methodology for extracting field-dependent intramolecular potentials from ab initio calculations is discussed, and the parameters controlling the energetics of intramolecular motion are directly coupled to the field experienced by the atoms in a molecule. Variability in the atomic charges is introduced via the fluctuating charge model of Rick et al. [S. Rick, et al., J. Chem. Phys. 1994, 101, (7), 6141-6156.]. Atomic charge fluctuations are much faster than atomic motion and, for practical reasons, a multiple time steps algorithm is required. In the implementation of MD simulations for this model, the Message Passing Interface (MPI) has been used. The simulation algorithm is complex for the fCINTRA model. However, with the help of load sharing, minimization of interprocessor communications, and code optimization, the overall simulation time is acceptable.
Intelligent software must frequently interrupt users, but the problem of deciding when to interrupt the user is still unsolved. In this paper, an algorithm is presented that identifies the appropriate times to interrupt the user is proposed and visualizations of the user’s state. The Intelligent Interruption Algorithm draws from user, task, and environmental contextual information dynamically extracted from the environment as the user performs computer based tasks. The visualizations are a representation of these complex components presented in a way that is meaningful for analysis purposes and for the user. This paper presents the following: the interruption algorithm; the machine learning algorithms used; and the preliminary results of visualizing the interruptible moment.
MPI is the predominant model for parallel programming in technical high performance computing. With an increasing number of cores and threads in cluster nodes the question arises whether pure MPI is an appropriate approach to utilize today’s compute clusters or if it is profitable to add another layer of parallelism within the nodes by applying OpenMP on a lower level. Investing a limited amount of manpower, we add OpenMP directives to three MPI production codes and compare and analyze the performance varying the number of MPI processes per node and the number of OpenMP threads per MPI process on current CMP/CMT architectures.
Gas hydrates represent both a bane and a potential boon to the oil and gas industry, and considerable research into hydrate formation has been undertaken. We have recently developed a multi-threaded version of a Monte Carlo crystal growth algorithm and applied it to simulate the growth of large structure II gas hydrates. This algorithm allows for the introduction of non-crystal molecules during simulations to study their effects on crystal growth rates and morphologies. Here, we report on our initial simulations of hydrate growth in the presence of poly(N-vinylpyrrolidone) (PVP). We have surveyed the PVP-hydrate interaction space by performing numerous simulations, each with a unique PVP-hydrate docking orientation. This survey produced a broad range of hydrate growth rates, ranging from almost complete suppression through to significant growth enhancement. The effect that PVP has on gas incorporation at crystal surfaces appears to be critical for differentiating between crystal growth inhibition and enhancement.
We present a parallelized contracted basis-iterative calculation of vibrational energy levels of CH$_5^+$ (a 12D calculation). We use Radau polyspherical coordinates and basis functions that are products of eigenfunctions of bend and stretch Hamiltonians. The basis functions have amplitude in all of the 120 equivalent minima. Many low-lying levels are well converged. A new parallelization scheme is presented.
We document and compare the performance of an OpenMP parallelized simulation code for a mixed-culture biofilm model on a desktop workstation with two quad core Xeon processors, and on SGI Altix Systems with single core and dual core Itanium processors. The underlying model is a parabolic system of highly non-linear partial differential equations, which is discretized in time using a non-local Mickens scheme, and in space using a standard finite difference method.
This paper presents our analysis of the failure behavior of large scale systems using the failure logs collected by Los Alamos National Laboratory on 22 of their computing clusters.We note that not all nodes show similar failure behavior in the systems. Our objective, therefore, was to arrive at an ordering of nodes to be incrementally (one by one) selected for duplication so as to achieve a target MTTF for the system after duplicating the least number of nodes. We arrived at a model for the fault coverage provided by duplicating each node and ordered the nodes according to coverage provided by each node. As compared to traditional approach of randomly choosing nodes for duplication, our model‐driven approach provides improvements ranging from 82
One key aspect for the successful utilization of grid environments is how to efficiently schedule distributed and parallel applications to these configurations. It is also desirable to make the tlymatching operation of available resources as transparent as possible to the user. These aspects are especially important for grid environments formed by heterogeneous multi-cluster machines. In this paper, we present an approach that considers both computer resources and communication links. The approach is based on a combination of ontology and fuzzy logic. The ontology paradigm is employed as a standard interface to accept users's requirements for desired resources. The fuzzy logic algorithms are used to compute parameters for matching based on dynamically monitored values of processor usage and communication. Experimental results indicate that the proposed approach is successful in terms of gathering dynamically more appropriate distributed resources and communication links in multi-cluster environments.
In this work, the lattice Boltzmann method (LBM) is verified for direct numerical simulation (DNS) of wall-bounded turbulent flows by simulating fully-developed turbulent channel flow and comparing the results to the spectral data of Moser et al. [3]. The turbulence statistics compared include: mean velocity and pressure profiles, Reynolds stress profiles, skewness and flatness factors, the turbulence kinetic-energy budget, and one-dimensional energy spectra. Additionally, a scalability test is performed for the implementation of the LBM parallelised with OpenMP for shared-memory architectures. The effect of the domain decomposition algorithm is studied by comparing the performance of a channel flow simulation decomposed with a naïve decomposition method to a case in which the decomposition is computed using the METIS library [4].
Multi-cores and multi-processors loosely coupled architectures are interesting commodity off-the-shelf architectures with which to build multi-cluster configurations as grid environments. However, such grid environments face challenges that must be circumvented. Examples include the heterogeneity of computational resources (e.g. operating systems, processors), of different programming paradigms and of network protocols. Transport protocols are especially important, because they can directly impact the execution of distributed engineering and scientific applications. In this paper, we present an empirical study of the SCTP, XTP and TCP as transport protocols with which to support applications in multi-cluster configurations. The environment includes several networks of workstations (NOWs) interconnected via a wide area network. Our experiments consider throughput and latency parameters against number of flows; the results show performance differences from the three transport protocols over a real distributed production configurations.
In this paper, we discuss the parallelization of power systems simulation using modern clusters consisting of multi-core compute nodes interconnected by a low latency interconnect. We describe an implementation strategy that exploits three types of parallelization that are amenable to the various types of inter-core “connectivity” commonly seen in such clusters. We also consider a set of design criteria that are important to Small-Medium Enterprises (SMEs) and how they affect implementation strategy. We report the results of some initial experiments done using components that will form a part of the prototype system under construction and then discuss the expected performance gains for the entire system based on manpower invested. We also identify several key factors in design for SMEs that differentiate their parallel systems needs from those of large organizations and research scientists.
